some concurrent changes Jan was making to the bufmgr. Here's an
updated version of the patch -- it should apply cleanly to CVS
HEAD and passes the regression tests.
This patch makes the following changes:
- remove the UnlockAndReleaseBuffer() and UnlockAndWriteBuffer()
macros, and replace uses of them with calls to the appropriate
functions.
- remove a bunch of #ifdef BMTRACE code: it is ugly & broken
(i.e. it doesn't compile)
- make BufferReplace() return a bool, not an int
- cleanup some logic in bufmgr.c; should be functionality
equivalent to the previous code, just cleaner now
- remove the BM_PRIVATE flag as it is unused
- improve a few comments, etc.
pghackers proposal of 8-Nov. All the existing cross-type comparison
operators (int2/int4/int8 and float4/float8) have appropriate support.
The original proposal of storing the right-hand-side datatype as part of
the primary key for pg_amop and pg_amproc got modified a bit in the event;
it is easier to store zero as the 'default' case and only store a nonzero
when the operator is actually cross-type. Along the way, remove the
long-since-defunct bigbox_ops operator class.
Remove the 'strategy map' code, which was a large amount of mechanism
that no longer had any use except reverse-mapping from procedure OID to
strategy number. Passing the strategy number to the index AM in the
first place is simpler and faster.
This is a preliminary step in planned support for cross-datatype index
operations. I'm committing it now since the ScanKeyEntryInitialize()
API change touches quite a lot of files, and I want to commit those
changes before the tree drifts under me.
protocol, per report from Igor Shevchenko. NOTIFY thought it could
do its thing if transaction blockState is TBLOCK_DEFAULT, but in
reality it had better check the low-level transaction state is
TRANS_DEFAULT as well. Formerly it was not possible to wait for the
client in a state where the first is true and the second is not ...
but now we can have such a state. Minor cleanup in StartTransaction()
as well.
discussion on pgsql-hackers: in READ COMMITTED mode we just have to force
a QuerySnapshot update in the trigger, but in SERIALIZABLE mode we have
to run the scan under a current snapshot and then complain if any rows
would be updated/deleted that are not visible in the transaction snapshot.
invalid (has the wrong magic number) until the build is entirely
complete. This turns out to cost no additional writes in the normal
case, since we were rewriting the metapage at the end of the process
anyway. In normal scenarios there's no real gain in security, because
a failed index build would roll back the transaction leaving an unused
index file, but for rebuilding shared system indexes this seems to add
some useful protection.
post-abort cleanup hooks. I'm surprised that we have not needed this
already, but I need it now to fix a plpgsql problem, and the usefulness
for other dynamically loaded modules seems obvious.
now able to cope with assigning new relfilenode values to nailed-in-cache
indexes, so they can be reindexed using the fully crash-safe method. This
leaves only shared system indexes as special cases. Remove the 'index
deactivation' code, since it provides no useful protection in the shared-
index case. Require reindexing of shared indexes to be done in standalone
mode, but remove other restrictions on REINDEX. -P (IgnoreSystemIndexes)
now prevents using indexes for lookups, but does not disable index updates.
It is therefore safe to allow from PGOPTIONS. Upshot: reindexing system catalogs
can be done without a standalone backend for all cases except
shared catalogs.
really general fix might be difficult, I believe the only case where
AtCommit_Notify could see an uncommitted tuple is where the other guy
has just unlistened and not yet committed. The best solution seems to
be to just skip updating that tuple, on the assumption that the other
guy does not want to hear about the notification anyway. This is not
perfect --- if the other guy rolls back his unlisten instead of committing,
then he really should have gotten this notify. But to do that, we'd have
to wait to see if he commits or not, or make UNLISTEN hold exclusive lock
on pg_listener until commit. Either of these answers is deadlock-prone,
not to mention horrible for interactive performance. Do it this way
for now. (What happened to that project to do LISTEN/NOTIFY in memory
with no table, anyway?)
pghackers. This fixes the problem recently reported by Markus KrÌutner
(hash bucket split corrupts the state of scans being done concurrently),
and I believe it also fixes all the known problems with deadlocks in
hash index operations. Hash indexes are still not really ready for prime
time (since they aren't WAL-logged), but this is a step forward.
killed items; just skip to the next item immediately. Only check for
key equality when we reach a non-killed item or the end of the index
page. This saves key comparisons when there are lots of killed items,
as for example in a heavily-updated table that's not been vacuumed lately.
Seems to be a win for pgbench anyway.
layout; therefore, this change forces REINDEX of hash indexes (though
not a full initdb). Widen hashm_ntuples to double so that hash space
management doesn't get confused by more than 4G entries; enlarge the
allowed number of free-space-bitmap pages; replace the useless bshift
field with a useful bmshift field; eliminate 4 bytes of wasted space
in the per-page special area.
scheme. A pleasant side effect is that it is *much* faster when deleting
a large fraction of the indexed tuples, because of elimination of
redundant hash_step activity induced by hash_adjscans. Various other
continuing code cleanup.
yet). Fix a couple of bugs that would only appear if multiple bitmap pages
are used, including a buffer reference leak and incorrect computation of bit
indexes. Get rid of 'overflow address' concept, which accomplished nothing
except obfuscating the code and creating a risk of failure due to limited
range of offset field. Rename some misleadingly-named fields and routines,
and improve documentation.
explanation of the remarkably confusing page addressing scheme.
The file also includes my planned-but-not-yet-implemented revision
of the hash index locking scheme.
target columns in INSERT and UPDATE targetlists. Don't rely on resname
to be accurate in ruleutils, either. This fixes bug reported by
Donald Fraser, in which renaming a column referenced in a rule did not
work very well.
index pages: when _bt_getbuf asks the FSM for a free index page, it is
possible (and, in some cases, even moderately likely) that the answer
will be the same page that _bt_split is trying to split. _bt_getbuf
already knew that the returned page might not be free, but it wasn't
prepared for the possibility that even trying to lock the page could
be problematic. Fix by doing a conditional rather than unconditional
grab of the page lock.
bottom. Otherwise we fail to moveright when the root page was split while
we were "in flight" to it. This is not a significant problem when the root
is above the leaf level, but if the root was also a leaf (ie, a single-page
index just got split) we may return the wrong leaf page to the caller,
resulting in failure to find a key that is in fact present. Bug has existed
at least since 7.1, probably forever.
fixed incorrect initial setting of StartUpID. The logic in XLogWrite()
expects that Write->curridx is advanced to the next page as soon as
LogwrtResult points to the end of the current page, but StartupXLOG()
failed to make that happen when the old WAL ended exactly on a page
boundary. Per trouble report from Hannu Krosing.
specific hash functions used by hash indexes, rather than the old
not-datatype-aware ComputeHashFunc routine. This makes it safe to do
hash joining on several datatypes that previously couldn't use hashing.
The sets of datatypes that are hash indexable and hash joinable are now
exactly the same, whereas before each had some that weren't in the other.
least-recently-used strategy from clog.c into slru.c. It doesn't
change any visible behaviour and passes all regression tests plus a
TruncateCLOG test done manually.
Apart from refactoring I made a little change to SlruRecentlyUsed,
formerly ClogRecentlyUsed: It now skips incrementing lru_counts, if
slotno is already the LRU slot, thus saving a few CPU cycles. To make
this work, lru_counts are initialised to 1 in SimpleLruInit.
SimpleLru will be used by pg_subtrans (part of the nested transactions
project), so the main purpose of this patch is to avoid future code
duplication.
Manfred Koizar
advertised in RowDescription message. Depending on the physical tuple's
column count is not really correct, since according to heap_getattr()
conventions the tuple may be short some columns, which will automatically
get read as nulls. Problem has been latent since forever, but was only
exposed by recent change to skip a projection step in SELECT * FROM...
example from Rao Kumar. This is a very corner corner-case, requiring
a minimum of three closely-spaced database crashes and an unlucky
positioning of the second recovery's checkpoint record before you'd notice
any problem. But the consequences are dire enough that it's a must-fix.
only remnant of this failed experiment is that the server will take
SET AUTOCOMMIT TO ON. Still TODO: provide some client-side autocommit
logic in libpq.
the type OID and typmod of the underlying base type. Per discussions
a few weeks ago with Andreas Pflug and others. Note that this behavioral
change affects both old- and new-protocol clients.
dead xlog segments are not considered part of a critical section. It is
not necessary to force a database-wide panic if we get a failure in these
operations. Per recent trouble reports.
handle multiple 'formats' for data I/O. Restructure CommandDest and
DestReceiver stuff one more time (it's finally starting to look a bit
clean though). Code now matches latest 3.0 protocol document as far
as message formats go --- but there is no support for binary I/O yet.
Neil Conway
Peter Eisentraut
Joe Conway
PostgreSQL Daemon
Tom Lane
Bruce Momjian