We had put a test for libxml2's xmlStructuredErrorContext variable in
configure, but of course that doesn't work on Windows builds.  The next
best alternative seems to be to test the LIBXML_VERSION symbol provided
by xmlversion.h.

Per report from Talha Bin Rizwan, though this fixes it in a different way
than his proposed patch.

In the initial cut at the "parameterized paths" feature, I'd simplified
create_index_paths() to the point where it would only generate a single
parameterized bitmap path per relation.  Experimentation with an example
supplied by Josh Berkus convinces me that that's not good enough: we really
need to consider a bitmap path for each possible outer relation.  Otherwise
we have regressions relative to pre-9.2 versions, in which the planner
picks a plain indexscan where it should have used a bitmap scan in queries
involving three or more tables.  Indeed, after fixing this, several queries
in the regression tests show improved plans as a result of using bitmap not
plain indexscans.

We use a hash table to track the parents of inner pages, but when inserting
to a leaf page, the caller of gistbufferinginserttuples() must pass a
correct block number of the leaf's parent page. Before gistProcessItup()
descends to a child page, it checks if the downlink needs to be adjusted to
accommodate the new tuple, and updates the downlink if necessary. However,
updating the downlink might require splitting the page, which might move the
downlink to a page to the right. gistProcessItup() doesn't realize that, so
when it descends to the leaf page, it might pass an out-of-date parent block
number as a result. Fix that by returning the block a tuple was inserted to
from gistbufferinginserttuples().

This fixes the bug reported by Zdeněk Jílovec.

The previous coding essentially assumed that nodes would be rescanned in
the same order they were initialized in; or at least that the "leader" of
a group of CTEscans would be rescanned before any others were required to
execute.  Unfortunately, that isn't even a little bit true.  It's possible
to devise queries in which the leader isn't rescanned until other CTEscans
on the same CTE have run to completion, or even in which the leader never
gets a rescan call at all.

The fix makes the leader specially responsible only for initial creation
and final destruction of the tuplestore; rescan resets are now a
symmetrically shared responsibility.  This means that we might reset the
tuplestore multiple times when restarting a plan subtree containing
multiple CTEscans; but resetting an already-empty tuplestore is cheap
enough that that doesn't seem like a problem.

Per report from Adam Mackler; the new regression test cases are based on
his example query.

Back-patch to 8.4 where CTE scans were introduced.

This situation creates a dependency loop that confuses pg_dump and probably
other things.  Moreover, since the mental model is that the extension
"contains" schemas it owns, but "is contained in" its extschema (even
though neither is strictly true), having both true at once is confusing for
people too.  So prevent the situation from being set up.

Reported and patched by Thom Brown.  Back-patch to 9.1 where extensions
were added.

This essentially reverts commit e54b10a6,
in which I'd decided that the "last ditch" join logic was useless.  The
folly of that is now exposed by a report from Pavel Stehule: although the
function should always find at least one join in a self-contained join
problem, it can still fail to do so in a sub-problem created by artificial
from_collapse_limit or join_collapse_limit constraints.  Adjust the
comments to describe this, and simplify the code a bit to match the new
coding of the earlier loop in the function.

I'm not terribly happy about this: I still subscribe to the opinion stated
in the previous commit message that the "last ditch" code can obscure logic
bugs elsewhere.  But the alternative seems to be to complicate the earlier
tests for does-this-relation-have-a-join-clause to the point where they can
tell whether the join clauses link outside the current join sub-problem.
And that looks messy, slow, and possibly a source of bugs in itself.
In any case, now is not the time to be inserting experimental code into
9.2, so let's just go back to the time-tested solution.

xml_parse() would attempt to fetch external files or URLs as needed to
resolve DTD and entity references in an XML value, thus allowing
unprivileged database users to attempt to fetch data with the privileges
of the database server.  While the external data wouldn't get returned
directly to the user, portions of it could be exposed in error messages
if the data didn't parse as valid XML; and in any case the mere ability
to check existence of a file might be useful to an attacker.

The ideal solution to this would still allow fetching of references that
are listed in the host system's XML catalogs, so that documents can be
validated according to installed DTDs.  However, doing that with the
available libxml2 APIs appears complex and error-prone, so we're not going
to risk it in a security patch that necessarily hasn't gotten wide review.
So this patch merely shuts off all access, causing any external fetch to
silently expand to an empty string.  A future patch may improve this.

In HEAD and 9.2, also suppress warnings about undefined entities, which
would otherwise occur as a result of not loading referenced DTDs.  Previous
branches don't show such warnings anyway, due to different error handling
arrangements.

Credit to Noah Misch for first reporting the problem, and for much work
towards a solution, though this simplistic approach was not his preference.
Also thanks to Daniel Veillard for consultation.

Security: CVE-2012-3489

This command generated new pg_depend entries linking the index to the
constraint and the constraint to the table, which match the entries made
when a unique or primary key constraint is built de novo.  However, it did
not bother to get rid of the entries linking the index directly to the
table.  We had considered the issue when the ADD CONSTRAINT USING INDEX
patch was written, and concluded that we didn't need to get rid of the
extra entries.  But this is wrong: ALTER COLUMN TYPE wasn't expecting such
redundant dependencies to exist, as reported by Hubert Depesz Lubaczewski.
On reflection it seems rather likely to break other things as well, since
there are many bits of code that crawl pg_depend for one purpose or
another, and most of them are pretty naive about what relationships they're
expecting to find.  Fortunately it's not that hard to get rid of the extra
dependency entries, so let's do that.

Back-patch to 9.1, where ALTER TABLE ADD CONSTRAINT USING INDEX was added.

Should be limited to the maximum number of connections excluding
autovacuum workers, not including.

Add similar check for max_wal_senders, which should never be higher than
max_connections.

This prevents spurious attempts to archive xlog files after promotion of
standby, a bug introduced by cascading replication patch in 9.2.

Fujii Masao, simplified and extended to cover streaming by Simon Riggs

Cascading replication copied the incoming file into pg_xlog but
didn't set path correctly, so the first attempt to open file failed
causing it to loop around and look for file in pg_xlog. So the
earlier coding worked, but accidentally rather than by design.

Spotted by Fujii Masao, fix by Fujii Masao and Simon Riggs

This was broken in commit ed0b409d,
which revised the GlobalTransactionData struct to not include the
associated PGPROC as its first member, but overlooked one place where
a cast was used in reliance on that equivalence.

The most effective way of fixing this seems to be to create a new function
that looks up the GlobalTransactionData struct given the XID, and make
both TwoPhaseGetDummyBackendId and TwoPhaseGetDummyProc rely on that.

Per report from Robert Ross.

Dave Kerr

DecodeInterval() failed to honor the "range" parameter (the special SQL
syntax for indicating which fields appear in the literal string) if the
time was signed.  This seems inappropriate, so make it work like the
not-signed case.  The inconsistency was introduced in my commit
f867339c, which as noted in its log message
was only really focused on making SQL-compliant literals work per spec.
Including a sign here is not per spec, but if we're going to allow it
then it's reasonable to expect it to work like the not-signed case.

Also, remove bogus setting of tmask, which caused subsequent processing to
think that what had been given was a timezone and not an hh:mm(:ss) field,
thus confusing checks for redundant fields.  This seems to be an aboriginal
mistake in Lockhart's commit 2cf16424.

Add regression test cases to illustrate the changed behaviors.

Back-patch as far as 8.4, where support for spec-compliant interval
literals was added.

Range problem reported and diagnosed by Amit Kapila, tmask problem by me.

As noted by Noah Misch, btree_xlog_delete_get_latestRemovedXid is
critically dependent on the assumption that it's examining a consistent
state of the database.  This was undocumented though, so the
seemingly-unrelated check for no active HS sessions might be thought to be
merely an optional optimization.  Improve comments, and add an explicit
check of reachedConsistency just to be sure.

This function returns InvalidTransactionId (thereby killing all HS
transactions) in several cases that are not nearly unlikely enough for my
taste.  This commit doesn't attempt to fix those deficiencies, just
document them.

Back-patch to 9.2, not from any real functional need but just to keep the
branches more closely synced to simplify possible future back-patching.

In yesterday's commit 962e0cc7, I added the
ResolveRecoveryConflictWithSnapshot call in the wrong place.  I correctly
put it before spgRedoVacuumRedirect itself would modify the index page ---
but not before RestoreBkpBlocks, so replay of a record with a full-page
image would modify the page before kicking off any conflicting HS
transactions.  Oops.

The correct test for whether a redirection tuple is removable is whether
tuple's xid < RecentGlobalXmin, not OldestXmin; the previous coding
failed to protect index searches being done in concurrent transactions that
have no XID.  This mirrors the recent fix in btree's page recycling logic
made in commit d3abbbeb.

Also, WAL-log the newest XID of any removed redirection tuple on an index
page, and apply ResolveRecoveryConflictWithSnapshot during InHotStandby WAL
replay.  This protects against concurrent Hot Standby transactions possibly
needing to see the redirection tuple(s).

Per my query of 2012-03-12 and subsequent discussion.

Parse analysis neglected to cover the case of a WITH clause attached to an
intermediate-level set operation; it only handled WITH at the top level
or WITH attached to a leaf-level SELECT.  Per report from Adam Mackler.

In HEAD, I rearranged the order of SelectStmt's fields to put withClause
with the other fields that can appear on non-leaf SelectStmts.  In back
branches, leave it alone to avoid a possible ABI break for third-party
code.

Back-patch to 8.4 where WITH support was added.

In the original coding of the log rotation stuff, we did not bother to make
the truncation logic work for the very first rotation after postmaster
start (or after a syslogger crash and restart).  It just always appended
in that case.  It did not seem terribly important at the time, but we've
recently had two separate complaints from people who expected it to work
unsurprisingly.  (Both users tend to restart the postmaster about as often
as a log rotation is configured to happen, which is maybe not typical use,
but still...)  Since the initial log file is opened in the postmaster,
fixing this requires passing down some more state to the syslogger child
process.

It's always been like this, so back-patch to all supported branches.

In the original coding of the autovacuum cancel feature, commit
acac68b2, an autovacuum process was
considered a target for cancellation if it was found to hard-block any
process examined in the deadlock search.  This patch tightens the test so
that the autovacuum must directly hard-block the current process.  This
should make the behavior more predictable in general, and in particular
it ensures that an autovacuum will not be canceled with less than
deadlock_timeout grace period.  In the old coding, it was possible for an
autovacuum to be canceled almost instantly, given unfortunate timing of two
or more other processes' lock attempts.

This also justifies the logging methodology in the recent commit
d7318d43; without this restriction, that
patch isn't providing enough information to see the connection of the
canceling process to the autovacuum.  Like that one, patch all the way
back.

The old message was at DEBUG2, so typically it didn't show up in the
log at all.  As a result, in most cases where autovacuum was canceled,
the only information that was logged was the table being vacuumed,
with no indication as to what problem caused the cancel.  Crank up
the level to LOG and add some more details to assist with debugging.

Back-patch all the way, per discussion on pgsql-hackers.

If a crash occurred immediately after the first nextval() call for a serial
column, WAL replay would restore the sequence to a state in which it
appeared that no nextval() had been done, thus allowing the first sequence
value to be returned again by the next nextval() call; as reported in
bug #6748 from Xiangming Mei.

More generally, the problem would occur if an ALTER SEQUENCE was executed
on a freshly created or reset sequence.  (The manifestation with serial
columns was introduced in 8.2 when we added an ALTER SEQUENCE OWNED BY step
to serial column creation.)  The cause is that sequence creation attempted
to save one WAL entry by writing out a WAL record that made it appear that
the first nextval() had already happened (viz, with is_called = true),
while marking the sequence's in-database state with log_cnt = 1 to show
that the first nextval() need not emit a WAL record.  However, ALTER
SEQUENCE would emit a new WAL entry reflecting the actual in-database state
(with is_called = false).  Then, nextval would allocate the first sequence
value and set is_called = true, but it would trust the log_cnt value and
not emit any WAL record.  A crash at this point would thus restore the
sequence to its post-ALTER state, causing the next nextval() call to return
the first sequence value again.

To fix, get rid of the idea of logging an is_called status different from
reality.  This means that the first nextval-driven WAL record will happen
at the first nextval call not the second, but the marginal cost of that is
pretty negligible.  In addition, make sure that ALTER SEQUENCE resets
log_cnt to zero in any case where it touches sequence parameters that
affect future nextval results.  This will result in some user-visible
changes in the contents of a sequence's log_cnt column, as reflected in the
patch's regression test changes; but no application should be depending on
that anyway, since it was already true that log_cnt changes rather
unpredictably depending on checkpoint timing.

In addition, make some basically-cosmetic improvements to get rid of
sequence.c's undesirable intimacy with page layout details.  It was always
really trying to WAL-log the contents of the sequence tuple, so we should
have it do that directly using a HeapTuple's t_data and t_len, rather than
backing into it with some magic assumptions about where the tuple would be
on the sequence's page.

Back-patch to all supported branches.

The initially implemented syntax, "CHECK NO INHERIT (expr)" was not
deemed very good, so switch to "CHECK (expr) NO INHERIT" instead.  This
way it looks similar to SQL-standards compliant constraint attribute.

Backport to 9.2 where the new syntax and feature was introduced.

Per discussion.

We made use of the ROWS estimate for set-returning functions used in FROM,
but not for those used in SELECT targetlists; which is a bit of an
oversight considering there are common usages that require the latter
approach.  Improve that.  (I had initially thought it might be worth
folding this into cost_qual_eval, but after investigation concluded that
that wouldn't be very helpful, so just do it separately.)  Per complaint
from David Johnston.

Back-patch to 9.2, but not further, for fear of destabilizing plan choices
in existing releases.

The code was setting it true for other constraints, which is
bogus.  Doing so caused bogus catalog entries for such constraints, and
in particular caused an error to be raised when trying to drop a
constraint of types other than CHECK from a table that has children,
such as reported in bug #6712.

In 9.2, additionally ignore connoinherit=true for other constraint
types, to avoid having to force initdb; existing databases might already
contain bogus catalog entries.

Includes a catversion bump (in HEAD only).

Bug report from Miroslav Šulc
Analysis from Amit Kapila and Noah Misch; Amit also contributed the patch.

When a whole-row Var is reading the result of a subquery, we need it to
ignore any "resjunk" columns that the subquery might have evaluated for
GROUP BY or ORDER BY purposes.  We've hacked this area before, in commit
68e40998, but that fix only covered
whole-row Vars of named composite types, not those of RECORD type; and it
was mighty klugy anyway, since it just assumed without checking that any
extra columns in the result must be resjunk.  A proper fix requires getting
hold of the subquery's targetlist so we can actually see which columns are
resjunk (whereupon we can use a JunkFilter to get rid of them).  So bite
the bullet and add some infrastructure to make that possible.

Per report from Andrew Dunstan and additional testing by Merlin Moncure.
Back-patch to all supported branches.  In 8.3, also back-patch commit
292176a1, which for some reason I had
not done at the time, but it's a prerequisite for this change.

Instead of having one hash table entry per relation/fork/segment, just have
one per relation, and use bitmapsets to represent which specific segments
need to be fsync'd.  This eliminates the need to scan the whole hash table
to implement FORGET_RELATION_FSYNC, which fixes the O(N^2) behavior
recently demonstrated by Jeff Janes for cases involving lots of TRUNCATE or
DROP TABLE operations during a single checkpoint cycle.  Per an idea from
Robert Haas.

(FORGET_DATABASE_FSYNC still sucks, but since dropping a database is a
pretty expensive operation anyway, we'll live with that.)

In passing, improve the delayed-unlink code: remove the pass over the list
in mdpreckpt, since it wasn't doing anything for us except supporting a
useless Assert in mdpostckpt, and fix mdpostckpt so that it will absorb
fsync requests every so often when clearing a large backlog of deletion
requests.

We were sending one per fork, but a little bit of refactoring allows us
to send just one request with forknum == InvalidForkNumber.  This not only
reduces pressure on the shared-memory request queue, but saves repeated
traversals of the checkpointer's hash table.

Functions like range_eq, range_before etc. are exposed at the SQL-level, but
they're also used internally by the GiST consistent support function. The
code sharing was done by a hack, TrickFunctionCall2, which relied on the
knowledge that all the functions used fn_extra the same way. This commit
splits the functions into internal versions that take a TypeCacheEntry as
argument, and thin wrappers to expose the functions at the SQL-level. The
internal versions can then be called directly and in a less hacky way from
the GiST consistent function.

This is just cosmetic, but backpatch to 9.2 anyway, to avoid having a
different version of this code in the 9.2 branch. That would make
backpatching fixes in this area more difficult.

Alexander Korotkov

ForwardFsyncRequest() supposed that it could only be called in regular
backends, which used to be true; but since the splitup of bgwriter and
checkpointer, it is also called in the bgwriter.  We do not want to count
such calls in pg_stat_bgwriter.buffers_backend statistics, so fix things
so that they aren't.

(It's worth noting here that this implies an alarmingly large increase in
the expected amount of cross-process fsync request traffic, which may well
mean that the process splitup was not such a hot idea.)

mdinit() was misusing IsBootstrapProcessingMode() to decide whether to
create an fsync pending-operations table in the current process.  This led
to creating a table not only in the startup and checkpointer processes as
intended, but also in the bgwriter process, not to mention other auxiliary
processes such as walwriter and walreceiver.  Creation of the table in the
bgwriter is fatal, because it absorbs fsync requests that should have gone
to the checkpointer; instead they just sit in bgwriter local memory and are
never acted on.  So writes performed by the bgwriter were not being fsync'd
which could result in data loss after an OS crash.  I think there is no
live bug with respect to walwriter and walreceiver because those never
perform any writes of shared buffers; but the potential is there for
future breakage in those processes too.

To fix, make AuxiliaryProcessMain() export the current process's
AuxProcType as a global variable, and then make mdinit() test directly for
the types of aux process that should have a pendingOpsTable.  Having done
that, we might as well also get rid of the random bool flags such as
am_walreceiver that some of the aux processes had grown.  (Note that we
could not have fixed the bug by examining those variables in mdinit(),
because it's called from BaseInit() which is run by AuxiliaryProcessMain()
before entering any of the process-type-specific code.)

Back-patch to 9.2, where the problem was introduced by the split-up of
bgwriter and checkpointer processes.  The bogus pendingOpsTable exists
in walwriter and walreceiver processes in earlier branches, but absent
any evidence that it causes actual problems there, I'll leave the older
branches alone.

In all branches back to 8.3, this patch fixes a questionable assumption in
CompactCheckpointerRequestQueue/CompactBgwriterRequestQueue that there are
no uninitialized pad bytes in the request queue structs.  This would only
cause trouble if (a) there were such pad bytes, which could happen in 8.4
and up if the compiler makes enum ForkNumber narrower than 32 bits, but
otherwise would require not-currently-planned changes in the widths of
other typedefs; and (b) the kernel has not uniformly initialized the
contents of shared memory to zeroes.  Still, it seems a tad risky, and we
can easily remove any risk by pre-zeroing the request array for ourselves.
In addition to that, we need to establish a coding rule that struct
RelFileNode can't contain any padding bytes, since such structs are copied
into the request array verbatim.  (There are other places that are assuming
this anyway, it turns out.)

In 9.1 and up, the risk was a bit larger because we were also effectively
assuming that struct RelFileNodeBackend contained no pad bytes, and with
fields of different types in there, that would be much easier to break.
However, there is no good reason to ever transmit fsync or delete requests
for temp files to the bgwriter/checkpointer, so we can revert the request
structs to plain RelFileNode, getting rid of the padding risk and saving
some marginal number of bytes and cycles in fsync queue manipulation while
we are at it.  The savings might be more than marginal during deletion of
a temp relation, because the old code transmitted an entirely useless but
nonetheless expensive-to-process ForgetRelationFsync request to the
background process, and also had the background process perform the file
deletion even though that can safely be done immediately.

In addition, make some cleanup of nearby comments and small improvements to
the code in CompactCheckpointerRequestQueue/CompactBgwriterRequestQueue.

Formerly, when trying to copy both indexes and comments, CREATE TABLE LIKE
had to pre-assign names to indexes that had comments, because it made up an
explicit CommentStmt command to apply the comment and so it had to know the
name for the index.  This creates bad interactions with other indexes, as
shown in bug #6734 from Daniele Varrazzo: the preassignment logic couldn't
take any other indexes into account so it could choose a conflicting name.

To fix, add a field to IndexStmt that allows it to carry a comment to be
assigned to the new index.  (This isn't a user-exposed feature of CREATE
INDEX, only an internal option.)  Now we don't need preassignment of index
names in any situation.

I also took the opportunity to refactor DefineIndex to accept the IndexStmt
as such, rather than passing all its fields individually in a mile-long
parameter list.

Back-patch to 9.2, but no further, because it seems too dangerous to change
IndexStmt or DefineIndex's API in released branches.  The bug exists back
to 9.0 where CREATE TABLE LIKE grew the ability to copy comments, but given
the lack of prior complaints we'll just let it go unfixed before 9.2.