Introduce latches. A latch is a boolean variable, with the capability to
wait until it is set. Latches can be used to reliably wait until a signal
arrives, which is hard otherwise because signals don't interrupt select()
on some platforms, and even when they do, there's race conditions.
On Unix, latches use the so called self-pipe trick under the covers to
implement the sleep until the latch is set, without race conditions. On
Windows, Windows events are used.
Use the new latch abstraction to sleep in walsender, so that as soon as
a transaction finishes, walsender is woken up to immediately send the WAL
to the standby. This reduces the latency between master and standby, which
is good.
Preliminary work by Fujii Masao. The latch implementation is by me, with
helpful comments from many people.
Tom Lane [Thu, 9 Sep 2010 17:19:40 +0000 (17:19 +0000)]
Remove obsolete claim that gzip is needed while installing PG's documentation.
It isn't, now that we ship the docs as loose files rather than a sub-tarball.
Also adjust the wording in a couple of places to make the lists of required
software read more consistently.
Bruce Momjian [Thu, 9 Sep 2010 00:48:22 +0000 (00:48 +0000)]
Doc fixes:
- remove excessive table cells
- moving function parameters into function tags rather than having
them being considered separate
- add return type column on XML2 contrib module functions list and
removing return types from function
- add table header to XML2 contrib parameter table
Bruce Momjian [Tue, 7 Sep 2010 14:10:30 +0000 (14:10 +0000)]
Modify pg_upgrade to set/restore all environment variables related to
collation/encoding to match English when reading controldata. This now
matches the English variable setting used by pg_regress.c.
Tom Lane [Sun, 5 Sep 2010 15:45:42 +0000 (15:45 +0000)]
GROUP BY can only infer functional dependency from non-deferrable primary keys.
Peter's original patch had this right, but I dropped the check while revising
the code to search pg_constraint instead of pg_index. Spotted by Dean Rasheed.
Tom Lane [Sat, 4 Sep 2010 17:45:56 +0000 (17:45 +0000)]
Pad the ps_status display with nulls, not blanks, on Darwin.
A long time ago, this didn't work nicely, but it seems to work on all recent
versions of OS X. The blank-pad method is less desirable since it results
in lots of extra space in ps' output. Per Alexey Klyukin.
Tom Lane [Fri, 3 Sep 2010 01:34:55 +0000 (01:34 +0000)]
Install a data-type-based solution for protecting pg_get_expr().
Since the code underlying pg_get_expr() is not secure against malformed
input, and can't practically be made so, we need to prevent miscreants
from feeding arbitrary data to it. We can do this securely by declaring
pg_get_expr() to take a new datatype "pg_node_tree" and declaring the
system catalog columns that hold nodeToString output to be of that type.
There is no way at SQL level to create a non-null value of type pg_node_tree.
Since the backend-internal operations that fill those catalog columns
operate below the SQL level, they are oblivious to the datatype relabeling
and don't need any changes.
Tom Lane [Thu, 2 Sep 2010 03:16:46 +0000 (03:16 +0000)]
Fix up flushing of composite-type typcache entries to be driven directly by
SI invalidation events, rather than indirectly through the relcache.
In the previous coding, we had to flush a composite-type typcache entry
whenever we discarded the corresponding relcache entry. This caused problems
at least when testing with RELCACHE_FORCE_RELEASE, as shown in recent report
from Jeff Davis, and might result in real-world problems given the kind of
unexpected relcache flush that that test mechanism is intended to model.
The new coding decouples relcache and typcache management, which is a good
thing anyway from a structural perspective. The cost is that we have to
search the typcache linearly to find entries that need to be flushed. There
are a couple of ways we could avoid that, but at the moment it's not clear
it's worth any extra trouble, because the typcache contains very few entries
in typical operation.
Back-patch to 8.2, the same as some other recent fixes in this general area.
The patch could be carried back to 8.0 with some additional work, but given
that it's only hypothetical whether we're fixing any problem observable in
the field, it doesn't seem worth the work now.
Tom Lane [Mon, 30 Aug 2010 16:46:23 +0000 (16:46 +0000)]
Fix oversight in RelFileNodeBackend patch: CreateFakeRelcacheEntry needs to
initialize the rd_backend field of a fake Relation entry correctly.
Fortunately, that is easy, since only non-temp relations should ever be
mentioned in the WAL stream.
Simon Riggs [Mon, 30 Aug 2010 14:16:48 +0000 (14:16 +0000)]
Teach GetOldestXmin() about KnownAssignedXids during recovery.
Very minor issue, though this is required for a later patch.
Reported by Heikki Linnakangas.
Tom Lane [Sun, 29 Aug 2010 19:33:14 +0000 (19:33 +0000)]
Reduce PANIC to ERROR in some occasionally-reported btree failure cases.
This patch changes _bt_split() and _bt_pagedel() to throw a plain ERROR,
rather than PANIC, for several cases that are reported from the field
from time to time:
* right sibling's left-link doesn't match;
* PageAddItem failure during _bt_split();
* parent page's next child isn't right sibling during _bt_pagedel().
In addition the error messages for these cases have been made a bit
more verbose, with additional values included.
The original motivation for PANIC here was to capture core dumps for
subsequent analysis. But with so many users whose platforms don't capture
core dumps by default, or who are unprepared to analyze them anyway, it's hard
to justify a forced database restart when we can fairly easily detect the
problems before we've reached the critical sections where PANIC would be
necessary. It is not currently known whether the reports of these messages
indicate well-hidden bugs in Postgres, or are a result of storage-level
malfeasance; the latter possibility suggests that we ought to try to be more
robust even if there is a bug here that's ultimately found.
Backpatch to 8.2. The code before that is sufficiently different that
it doesn't seem worth the trouble to back-port further.
Robert Haas [Fri, 27 Aug 2010 21:31:19 +0000 (21:31 +0000)]
Insert additional compiler placation into objectaddress.c.
Peter Eisentraut reports that some bits of the "address" variable
in get_object_address() give "may be used uninitialized" warnings;
this likes the only excuse his compiler could have for thinking
that's possible.
Tom Lane [Thu, 26 Aug 2010 22:00:19 +0000 (22:00 +0000)]
Document the existence of the socket lock file under unix_socket_directory,
which is perhaps not a terribly good spot for it but there doesn't seem to be
a better place. Also add a source-code comment pointing out a couple reasons
for having a separate lock file. Per suggestion from Greg Smith.
Tom Lane [Thu, 26 Aug 2010 19:58:36 +0000 (19:58 +0000)]
Update time zone data files to tzdata release 2010l: DST law changes in
Egypt and Palestine. Added new names for two Micronesian timezones:
Pacific/Chuuk is now preferred over Pacific/Truk (and the preferred
abbreviation is CHUT not TRUT) and Pacific/Pohnpei is preferred over
Pacific/Ponape. Historical corrections for Finland.
Alvaro Herrera [Thu, 26 Aug 2010 19:49:08 +0000 (19:49 +0000)]
Improve wording for privilege description on certain failure messages; the
original misleadingly suggests that only access is meant, causing confusion.
Per recent trouble report by Robert McGehee on pgsql-admin.
Tom Lane [Thu, 26 Aug 2010 18:54:37 +0000 (18:54 +0000)]
Fix ExecMakeTableFunctionResult to verify that all rows returned by a SRF
returning "record" actually do have the same rowtype. This is needed because
the parser can't realistically enforce that they will all have the same typmod,
as seen in a recent example from David Wheeler.
Back-patch to 8.0, which is as far back as we have the notion of RECORD
subtypes being distinguished by typmod. Wheeler's example depends on
8.4-and-up features, but I suspect there may be ways to provoke similar
failures before 8.4.
Tom Lane [Thu, 26 Aug 2010 18:34:37 +0000 (18:34 +0000)]
Don't auto-create the subdirectories holding built documentation in a VPATH
build tree. If we actually build the docs in the VPATH tree, those dirs
will get created then; but if they're present and empty, they capture the
vpathsearch searches in "make install", preventing installation of prebuilt
docs that might exist in the source tree. Per bug #5595 from Dmtiriy Igrishin.
Fix based on idea from Peter Eisentraut.
Tom Lane [Wed, 25 Aug 2010 21:42:55 +0000 (21:42 +0000)]
Document filtering dictionaries in textsearch.sgml.
While at it, copy-edit the description of prefix-match marker support in
synonym dictionaries, and clarify the description of the default unaccent
dictionary a bit more.
Tom Lane [Wed, 25 Aug 2010 20:10:55 +0000 (20:10 +0000)]
Improve hint message for ENOMEM failure from shmget().
It turns out that some platforms return ENOMEM for a request that violates
SHMALL, whereas we were assuming that ENOSPC would always be used for that.
Apparently the latter is a Linuxism while ENOMEM is the BSD tradition.
Extend the ENOMEM hint to suggest that raising SHMALL might be needed.
Per gripe from A.M.
Backpatch to 9.0, but not further, because this doesn't seem important
enough to warrant creating extra translation work in the stable branches.
(If it were, we'd have figured this out years ago.)
Peter Eisentraut [Wed, 25 Aug 2010 19:37:56 +0000 (19:37 +0000)]
Catch null pointer returns from PyCObject_AsVoidPtr and PyCObject_FromVoidPtr
This is reproducibly possible in Python 2.7 if the user turned
PendingDeprecationWarning into an error, but it's theoretically also possible
in earlier versions in case of exceptional conditions.
Peter Eisentraut [Tue, 24 Aug 2010 18:06:12 +0000 (18:06 +0000)]
When in automatic dependency mode, never delete any intermediate
files automatically. Otherwise, the following could happen: When
starting from a clean source tree, the first build would delete the
intermediate file, but also create the dependency file, which
mentions the intermediate file, thus making it non-intermediate.
The second build will then need to rebuild the now non-intermediate
missing file. So the second build will do work even though nothing
had changed. One place where this happens is the .c -> .o -> .so
chain for some contrib modules.
Tom Lane [Mon, 23 Aug 2010 17:20:01 +0000 (17:20 +0000)]
Marginal code cleanup for streaming replication.
There is no reason that proc.c should have to get involved in this dirty hack
for letting the postmaster know which children are walsenders. Revert that
file to the way it was, and confine the kluge to pmsignal.c and postmaster.c.
Tom Lane [Sat, 21 Aug 2010 16:55:51 +0000 (16:55 +0000)]
Use a non-locale-dependent definition of isspace() in array_in/array_out.
array_in discards unquoted leading and trailing whitespace in array values,
while array_out is careful to quote array elements that contain whitespace.
This is problematic when the definition of "whitespace" varies between
locales: array_in could drop characters that were meant to be part of the
value. To avoid that, lock down "whitespace" to mean only the traditional
six ASCII space characters.
This change also works around a bug in OS X and some older BSD systems, in
which isspace() could return true for character fragments in UTF8 locales.
(There may be other places in PG where that bug could cause problems, but
this is the only one complained of so far; see recent report from Steven
Schlansker.)
Back-patch to 9.0, but not further. Given the lack of previous reports
of trouble, changing this behavior in stable branches seems to offer
more risk of breaking applications than reward of avoiding problems.
Tom Lane [Sat, 21 Aug 2010 13:59:44 +0000 (13:59 +0000)]
Improve parallel restore's ability to cope with selective restore (-L option).
The original coding tended to break down in the face of modified restore
orders, as shown in bug #5626 from Albert Ullrich, because it would flip over
into parallel-restore operation too soon. That causes problems because we
don't have sufficient dependency information in dump archives to allow safe
parallel processing of SECTION_PRE_DATA items. Even if we did, it's probably
undesirable to allow that to override the commanded restore order.
To fix the problem of omitted items causing unexpected changes in restore
order, tweak SortTocFromFile so that omitted items end up at the head of
the list not the tail. This ensures that they'll be examined and their
dependencies will be marked satisfied before we get to any interesting
items.
In HEAD and 9.0, we can easily change restore_toc_entries_parallel so that
all SECTION_PRE_DATA items are guaranteed to be processed in the initial
serial-restore loop, and hence in commanded order. Only DATA and POST_DATA
items are candidates for parallel processing. For them there might be
variations from the commanded order because of parallelism, but we should
do it in a safe order thanks to dependencies.
In 8.4 it's much harder to make such a guarantee. I settled for not
letting the initial loop break out into parallel processing mode if
it sees a DATA/POST_DATA item that's not to be restored; this at least
prevents a non-restorable item from causing premature exit from the loop.
This means that 8.4 will be more likely to fail given a badly-ordered -L
list than 9.x, but we don't really promise any such thing will work anyway.
Robert Haas [Fri, 20 Aug 2010 01:07:50 +0000 (01:07 +0000)]
Remove the isLocalBuf argument from ReadBuffer_common.
Since an SMgrRelation now knows whether or not the underlying relation is
temporary, there's no point in also passing that information via an
additional argument.
Tom Lane [Thu, 19 Aug 2010 22:55:01 +0000 (22:55 +0000)]
Bring some sanity to the trace_recovery_messages code and docs.
Per gripe from Fujii Masao, though this is not exactly his proposed patch.
Categorize as DEVELOPER_OPTIONS and set context PGC_SIGHUP, as per Fujii,
but set the default to LOG because higher values aren't really sensible
(see the code for trace_recovery()). Fix the documentation to agree with
the code and to try to explain what the variable actually does. Get rid
of no-op calls trace_recovery(LOG), which accomplish nothing except to
demonstrate that this option confuses even its author.
Tom Lane [Thu, 19 Aug 2010 18:57:57 +0000 (18:57 +0000)]
Allow USING and INTO clauses of plpgsql's EXECUTE to appear in either order.
Aside from being more forgiving, this prevents a rather surprising misbehavior
when the "wrong" order was used: the old code didn't throw a syntax error,
but absorbed the INTO clause into the last USING expression, which then did
strange things downstream.
Intentionally not changing the documentation; we'll continue to advertise
only the "standard" clause order.
Backpatch to 8.4, where the USING clause was added to EXECUTE.
Tom Lane [Thu, 19 Aug 2010 18:10:48 +0000 (18:10 +0000)]
Keep exec_simple_check_plan() from thinking "SELECT foo INTO bar" is simple.
It's not clear if this situation can occur in plpgsql other than via the
EXECUTE USING case Heikki illustrated, which I will shortly close off.
However, ignoring the intoClause if it's there is surely wrong, so let's
patch it for safety.
Backpatch to 8.3, which is as far back as this code has a PlannedStmt
to deal with. There might be another way to make an equivalent test
before that, but since this is just preventing hypothetical bugs,
I'm not going to obsess about it.
Revert patch to coerce 'unknown' type parameters in the backend. As Tom
pointed out, it would need a 2nd pass after the whole query is processed to
correctly check that an unknown Param is coerced to the same target type
everywhere. Adding the 2nd pass would add a lot more code, which doesn't
seem worth the risk given that there isn't much of a use case for passing
unknown Params in the first place. The code would work without that check,
but it might be confusing and the behavior would be different from the
varparams case.
Instead, just coerce all unknown params in a PL/pgSQL USING clause to text.
That's simple, and is usually what users expect.
Revert the patch in CVS HEAD and master, and backpatch the new solution to
8.4. Unlike the previous solution, this applies easily to 8.4 too.
Tom Lane [Thu, 19 Aug 2010 16:16:20 +0000 (16:16 +0000)]
Allocate local buffers in a context of their own, rather than dumping them
into TopMemoryContext. This makes no functional difference, but makes it
easier to see what the space is being used for in MemoryContextStats dumps.
Per a recent example in which I was surprised by the size of TopMemoryContext.
Tom Lane [Thu, 19 Aug 2010 15:46:18 +0000 (15:46 +0000)]
Fix possible corruption of AfterTriggerEventLists in subtransaction rollback.
afterTriggerInvokeEvents failed to adjust events->tailfree when truncating
the last chunk of an event list. This could result in the data being
"de-truncated" by afterTriggerRestoreEventList during a subsequent
subtransaction abort. Even that wouldn't kill us, because the re-added data
would just be events marked DONE --- unless the data had been partially
overwritten by new events. Then we might crash, or in any case misbehave
(perhaps fire triggers twice, or fire triggers with the wrong event data).
Per bug #5622 from Thue Janus Kristensen.
Back-patch to 8.4 where the current trigger list representation was introduced.
Robert Haas [Thu, 19 Aug 2010 02:58:37 +0000 (02:58 +0000)]
Tidy up a few calls to smrgextend().
In the new API introduced by my patch to include the backend ID in
temprel filenames, the last argument to smrgextend() became skipFsync
rather than isTemp, but these calls didn't get the memo. It's not
really a problem to pass rel->rd_istemp rather than just plain false,
because smgrextend() now automatically skips the fsync for temprels
anyway, but this seems cleaner and saves some minute number of cycles.
Tom Lane [Wed, 18 Aug 2010 21:52:24 +0000 (21:52 +0000)]
Reset the per-output-tuple exprcontext each time through the main loop in
ExecModifyTable(). This avoids memory leakage when trigger functions leave
junk behind in that context (as they more or less must). Problem and solution
identified by Dean Rasheed.
I'm a bit concerned about the longevity of this solution --- once a plan can
have multiple ModifyTable nodes, we are very possibly going to have to do
something different. But it should hold up for 9.0.
Tom Lane [Wed, 18 Aug 2010 19:54:01 +0000 (19:54 +0000)]
Rename utf2ucs() to utf8_to_unicode(), and export it so it can be used
elsewhere.
Similarly rename the version in mbprint.c, not because this affects anything
but just to keep the two copies in exact sync. There was some discussion of
having only one copy in src/port/ instead, but this function is so small
and unlikely to change that that seems like overkill.
Slightly editorialized version of a patch by Joseph Adams. (The bug-fix
aspect of his patch was applied separately, and back-patched.)
Tom Lane [Wed, 18 Aug 2010 18:35:21 +0000 (18:35 +0000)]
Fix failure of "ALTER TABLE t ADD COLUMN c serial" when done by non-owner.
The implicitly created sequence was created as owned by the current user,
who could be different from the table owner, eg if current user is a
superuser or some member of the table's owning role. This caused sanity
checks in the SEQUENCE OWNED BY code to spit up. Although possibly we
don't need those sanity checks, the safest fix seems to be to make sure
the implicit sequence is assigned the same owner role as the table has.
(We still do all permissions checks as the current user, however.)
Per report from Josh Berkus.
Back-patch to 9.0. The bug goes back to the invention of SEQUENCE OWNED BY
in 8.2, but the fix requires an API change for DefineRelation(), which seems
to have potential for breaking third-party code if done in a minor release.
Given the lack of prior complaints, it's probably not worth fixing in the
stable branches.
Tom Lane [Wed, 18 Aug 2010 15:21:54 +0000 (15:21 +0000)]
Add missing handling of PlannedStmt.transientPlan in copyfuncs/outfuncs.
_outPlannedStmt is only debug support, so the omission there was not very
serious, but the omission in _copyPlannedStmt is a real bug. The consequence
would be that a copied plan tree would never be marked as a transient plan,
so that we would forget we ought to replan it after some not-yet-ready index
becomes ready for use. This might explain some past complaints about indexes
created with CREATE INDEX CONCURRENTLY not being used right away. Problem
spotted by Yeb Havinga.
Coerce 'unknown' type parameters to the right type in the fixed-params
parse_analyze() function. That case occurs e.g with PL/pgSQL
EXECUTE ... USING 'stringconstant'.
The coercion with a CoerceViaIO node. The result is similar to the coercion
via input function performed for unknown constants in coerce_type(),
except that this happens at runtime.
Backpatch to 9.0. The issue is present in 8.4 as well, but the coerce param
hook infrastructure this patch relies on was introduced in 9.0. Given the
lack of user reports and harmlessness of the bug, it's not worth attempting
a different fix just for 8.4.
Tom Lane [Mon, 16 Aug 2010 17:32:46 +0000 (17:32 +0000)]
Arrange to fsync the contents of lockfiles (both postmaster.pid and the
socket lockfile) when writing them. The lack of an fsync here may well
explain two different reports we've seen of corrupted lockfile contents,
which doesn't particularly bother the running server but can prevent a
new server from starting if the old one crashes. Per suggestion from
Alvaro.
Robert Haas [Mon, 16 Aug 2010 02:02:28 +0000 (02:02 +0000)]
Make LockDatabaseObject() AcceptInvalidationMessages().
This is appropriate for the same reasons we already do it in
LockSharedObject(): things might have changed while we were waiting
for the lock. There doesn't seem to be a live bug here at the moment,
but that's mostly because it isn't currently used for very much.
Tom Lane [Mon, 16 Aug 2010 00:06:18 +0000 (00:06 +0000)]
Fix psql's copy of utf2ucs() to match the backend's copy exactly;
in particular, propagate a fix in the test to see whether a UTF8 character has
length 4 bytes. This is likely of little real-world consequence because
5-or-more-byte UTF8 sequences are not supported by Postgres nor seen anywhere
in the wild, but still we may as well get it right. Problem found by Joseph
Adams.
Tom Lane [Sat, 14 Aug 2010 15:47:13 +0000 (15:47 +0000)]
Fix planner to make a reasonable assumption about the amount of memory space
used by array_agg(), string_agg(), and similar aggregate functions that use
"internal" as their transition datatype. The previous coding thought this
took *no* extra space, since "internal" is pass-by-value; but actually these
aggregates typically consume a great deal of space. Per bug #5608 from
Itagaki Takahiro, and fix suggestion from Hitoshi Harada.
Back-patch to 8.4, where array_agg was introduced.
Robert Haas [Fri, 13 Aug 2010 20:10:54 +0000 (20:10 +0000)]
Include the backend ID in the relpath of temporary relations.
This allows us to reliably remove all leftover temporary relation
files on cluster startup without reference to system catalogs or WAL;
therefore, we no longer include temporary relations in XLOG_XACT_COMMIT
and XLOG_XACT_ABORT WAL records.
Since these changes require including a backend ID in each
SharedInvalSmgrMsg, the size of the SharedInvalidationMessage.id
field has been reduced from two bytes to one, and the maximum number
of connections has been reduced from INT_MAX / 4 to 2^23-1. It would
be possible to remove these restrictions by increasing the size of
SharedInvalidationMessage by 4 bytes, but right now that doesn't seem
like a good trade-off.
Tom Lane [Fri, 13 Aug 2010 20:04:33 +0000 (20:04 +0000)]
Minor #include cleanup.
I just noticed that libpq's pqsignal.h was violating our general inclusion
style guidelines by explicitly including postgres_fe.h. Remove that, and
put it in pqsignal.c where it belongs.
Tom Lane [Fri, 13 Aug 2010 18:36:26 +0000 (18:36 +0000)]
Add xml_is_well_formed, xml_is_well_formed_document, xml_is_well_formed_content
functions to the core XML code. Per discussion, the former depends on
XMLOPTION while the others do not. These supersede a version previously
offered by contrib/xml2.