Tom Lane [Fri, 14 Jun 2019 15:25:13 +0000 (11:25 -0400)]
Attempt to identify system timezone by reading /etc/localtime symlink.
On many modern platforms, /etc/localtime is a symlink to a file within the
IANA database. Reading the symlink lets us find out the name of the system
timezone directly, without going through the brute-force search embodied in
scan_available_timezones(). This shortens the runtime of initdb by some
tens of ms, which is helpful for the buildfarm, and it also allows us to
reliably select the same zone name the system was actually configured for,
rather than possibly choosing one of IANA's many zone aliases. (For
example, in a system configured for "Asia/Tokyo", the brute-force search
would not choose that name but its alias "Japan", on the grounds of the
latter string being shorter. More surprisingly, "Navajo" is preferred
to either "America/Denver" or "US/Mountain", as seen in an old complaint
from Josh Berkus.)
If /etc/localtime doesn't exist, or isn't a symlink, or we can't make
sense of its contents, or the contents match a zone we know but that
zone doesn't match the observed behavior of localtime(), fall back to
the brute-force search.
Also, tweak initdb so that it prints the zone name it selected.
In passing, replace the last few references to the "Olson" database in
code comments with "IANA", as that's been our preferred term since
commit b2cbced9e.
Back-patch of commit 23bd3cec6. The original intention was to not
back-patch, since this can result in cosmetic behavioral changes ---
for example, on my own workstation initdb now chooses "America/New_York",
where it used to prefer "US/Eastern" which is equivalent and shorter.
However, our hand has been more or less forced by tzdb update 2019a,
which made the "UCT" zone fully equivalent to "UTC". Our old code
now prefers "UCT" on the grounds of it being alphabetically first,
and that's making nobody happy. Choosing the alias indicated by
/etc/localtime is a more defensible behavior. (Users who don't like
the results can always force the decision by setting the TZ environment
variable before running initdb.)
Patch by me, per a suggestion from Robert Haas; review by Michael Paquier
Alvaro Herrera [Thu, 13 Jun 2019 21:28:24 +0000 (17:28 -0400)]
Avoid spurious deadlocks when upgrading a tuple lock
When two (or more) transactions are waiting for transaction T1 to release a
tuple-level lock, and transaction T1 upgrades its lock to a higher level, a
spurious deadlock can be reported among the waiting transactions when T1
finishes. The simplest example case seems to be:
T1: select id from job where name = 'a' for key share;
Y: select id from job where name = 'a' for update; -- starts waiting for X
Z: select id from job where name = 'a' for key share;
T1: update job set name = 'b' where id = 1;
Z: update job set name = 'c' where id = 1; -- starts waiting for X
T1: rollback;
At this point, transaction Y is rolled back on account of a deadlock: Y
holds the heavyweight tuple lock and is waiting for the Xmax to be released,
while Z holds part of the multixact and tries to acquire the heavyweight
lock (per protocol) and goes to sleep; once X releases its part of the
multixact, Z is awakened only to be put back to sleep on the heavyweight
lock that Y is holding while sleeping. Kaboom.
This can be avoided by having Z skip the heavyweight lock acquisition. As
far as I can see, the biggest downside is that if there are multiple Z
transactions, the order in which they resume after X finishes is not
guaranteed.
Backpatch to 9.6. The patch applies cleanly on 9.5, but the new tests don't
work there (because isolationtester is not smart enough), so I'm not going
to risk it.
Tom Lane [Thu, 13 Jun 2019 14:53:17 +0000 (10:53 -0400)]
Mark ReplicationSlotCtl as PGDLLIMPORT.
Also MyReplicationSlot, in branches where it wasn't already.
This was discussed in the thread that resulted in c572599c6, but
for some reason nobody pulled the trigger. Now that we have another
request for the same thing, we should just do it.
Etsuro Fujita [Thu, 13 Jun 2019 08:59:11 +0000 (17:59 +0900)]
postgres_fdw: Account for triggers in non-direct remote UPDATE planning.
Previously, in postgresPlanForeignModify, we planned an UPDATE operation
on a foreign table so that we transmit only columns that were explicitly
targets of the UPDATE, so as to avoid unnecessary data transmission, but
if there were BEFORE ROW UPDATE triggers on the foreign table, those
triggers might change values for non-target columns, in which case we
would miss sending changed values for those columns. Prevent optimizing
away transmitting all columns if there are BEFORE ROW UPDATE triggers on
the foreign table.
This is an oversight in commit 7cbe57c34 which added triggers on foreign
tables, so apply the patch all the way back to 9.4 where that came in.
Author: Shohei Mochizuki Reviewed-by: Amit Langote
Discussion: https://postgr.es/m/201905270152.x4R1q3qi014550@toshiba.co.jp
Tom Lane [Thu, 13 Jun 2019 02:54:46 +0000 (22:54 -0400)]
Doc: improve description of allowed spellings for Boolean input.
datatype.sgml failed to explain that boolin() accepts any unique
prefix of the basic input strings. Indeed it was actively misleading
because it called out a few minimal prefixes without mentioning that
there were more valid inputs.
I also felt that it wasn't doing anybody any favors by conflating
SQL key words, valid Boolean input, and string literals containing
valid Boolean input. Rewrite in hopes of reducing the confusion.
Per bug #15836 from Yuming Wang, as diagnosed by David Johnston.
Back-patch to supported branches.
Tom Lane [Wed, 12 Jun 2019 23:42:38 +0000 (19:42 -0400)]
Fix incorrect printing of queries with duplicated join names.
Given a query in which multiple JOIN nodes used the same alias
(which'd necessarily be in different sub-SELECTs), ruleutils.c
would assign the JOIN nodes distinct aliases for clarity ...
but then it forgot to print the modified aliases when dumping
the JOIN nodes themselves. This results in a dump/reload hazard
for views, because the emitted query is flat-out incorrect:
Vars will be printed with table names that have no referent.
This has been wrong for a long time, so back-patch to all supported
branches.
Tom Lane [Wed, 12 Jun 2019 21:29:48 +0000 (17:29 -0400)]
In walreceiver, don't try to do ereport() in a signal handler.
This is quite unsafe, even for the case of ereport(FATAL) where we won't
return control to the interrupted code, and despite this code's use of
a flag to restrict the areas where we'd try to do it. It's possible
for example that we interrupt malloc or free while that's holding a lock
that's meant to protect against cross-thread interference. Then, any
attempt to do malloc or free within ereport() will result in a deadlock,
preventing the walreceiver process from exiting in response to SIGTERM.
We hypothesize that this explains some hard-to-reproduce failures seen
in the buildfarm.
Hence, get rid of the immediate-exit code in WalRcvShutdownHandler,
as well as the logic associated with WalRcvImmediateInterruptOK.
Instead, we need to take care that potentially-blocking operations
in the walreceiver's data transmission logic (libpqwalreceiver.c)
will respond reasonably promptly to the process's latch becoming
set and then call ProcessWalRcvInterrupts. Much of the needed code
for that was already present in libpqwalreceiver.c. I refactored
things a bit so that all the uses of PQgetResult use latch-aware
waiting, but didn't need to do much more.
These changes should be enough to ensure that libpqwalreceiver.c
will respond promptly to SIGTERM whenever it's waiting to receive
data. In principle, it could block for a long time while waiting
to send data too, and this patch does nothing to guard against that.
I think that that hazard is mostly theoretical though: such blocking
should occur only if we fill the kernel's data transmission buffers,
and we don't generally send enough data to make that happen without
waiting for input. If we find out that the hazard isn't just
theoretical, we could fix it by using PQsetnonblocking, but that
would require more ticklish changes than I care to make now.
Back-patch of commit a1a789eb5. This problem goes all the way back
to the origins of walreceiver; but given the substantial reworking
the module received during the v10 cycle, it seems unsafe to assume
that our testing on HEAD validates this patch for pre-v10 branches.
And we'd need to back-patch some prerequisite patches (at least 597a87ccc and its followups, maybe other things), increasing the risk
of problems. Given the dearth of field reports matching this problem,
it's not worth much risk. Hence back-patch to v10 and v11 only.
Tom Lane [Wed, 12 Jun 2019 16:29:24 +0000 (12:29 -0400)]
Fix ALTER COLUMN TYPE failure with a partial exclusion constraint.
ATExecAlterColumnType failed to consider the possibility that an index
that needs to be rebuilt might be a child of a constraint that needs to be
rebuilt. We missed this so far because usually a constraint index doesn't
have a direct dependency on its table, just on the constraint object.
But if there's a WHERE clause, then dependency analysis of the WHERE
clause results in direct dependencies on the column(s) mentioned in WHERE.
This led to trying to drop and rebuild both the constraint and its
underlying index.
In v11/HEAD, we successfully drop both the index and the constraint,
and then try to rebuild both, and of course the second rebuild hits a
duplicate-index-name problem. Before v11, it fails with obscure messages
about a missing relation OID, due to trying to drop the index twice.
This is essentially the same kind of problem noted in commit 20bef2c31: the possible dependency linkages are broader than what
ATExecAlterColumnType was designed for. It was probably OK when
written, but it's certainly been broken since the introduction of
partial exclusion constraints. Fix by adding an explicit check
for whether any of the indexes-to-be-rebuilt belong to any of the
constraints-to-be-rebuilt, and ignoring any that do.
In passing, fix a latent bug introduced by commit 8b08f7d48: in
get_constraint_index() we must "continue" not "break" when rejecting
a relation of a wrong relkind. This is harmless today because we don't
expect that code path to be taken anyway; but if there ever were any
relations to be ignored, the existing coding would have an extremely
undesirable dependency on the order of pg_depend entries.
Also adjust a couple of obsolete comments.
Per bug #15835 from Yaroslav Schekin. Back-patch to all supported
branches.
Michael Paquier [Wed, 12 Jun 2019 02:30:41 +0000 (11:30 +0900)]
Fix handling of COMMENT for domain constraints
For a non-superuser, changing a comment on a domain constraint was
leading to a cache lookup failure as the code tried to perform the
ownership lookup on the constraint OID itself, thinking that it was a
type, but this check needs to happen on the type the domain constraint
relies on. As the type a domain constraint relies on can be guessed
directly based on the constraint OID, first fetch its type OID and
perform the ownership on it.
This is broken since 7eca575, which has split the handling of comments
for table constraints and domain constraints, so back-patch down to
9.5.
Reported-by: Clemens Ladisch
Author: Daniel Gustafsson, Michael Paquier Reviewed-by: Álvaro Herrera
Discussion: https://postgr.es/m/15833-808e11904835d26f@postgresql.org
Backpatch-through: 9.5
David Rowley [Tue, 11 Jun 2019 20:09:11 +0000 (08:09 +1200)]
doc: Add best practises section to partitioning docs
A few questionable partitioning designs have been cropping up lately
around the mailing lists. Generally, these cases have been partitioning
using too many partitions which have caused performance or OOM problems for
the users.
Since we have very little else to guide users into good design, here we
add a new section to the partitioning documentation with some best
practise guidelines for good design.
Tom Lane [Tue, 11 Jun 2019 17:33:08 +0000 (13:33 -0400)]
Fix conversion of JSON strings to JSON output columns in json_to_record().
json_to_record(), when an output column is declared as type json or jsonb,
should emit the corresponding field of the input JSON object. But it got
this slightly wrong when the field is just a string literal: it failed to
escape the contents of the string. That typically resulted in syntax
errors if the string contained any double quotes or backslashes.
jsonb_to_record() handles such cases correctly, but I added corresponding
test cases for it too, to prevent future backsliding.
Improve the documentation, as it provided only a very hand-wavy
description of the conversion rules used by these functions.
Per bug report from Robert Vollmert. Back-patch to v10 where the
error was introduced (by commit cf35346e8).
Note that PG 9.4 - 9.6 also get this case wrong, but differently so:
they feed the de-escaped contents of the string literal to json[b]_in.
That behavior is less obviously wrong, so possibly it's being depended on
in the field, so I won't risk trying to make the older branches behave
like the newer ones.
Andres Freund [Tue, 11 Jun 2019 06:20:48 +0000 (23:20 -0700)]
Don't access catalogs to validate GUCs when not connected to a DB.
Vignesh found this bug in the check function for
default_table_access_method's check hook, but that was just copied
from older GUCs. Investigation by Michael and me then found the bug in
further places.
When not connected to a database (e.g. in a walsender connection), we
cannot perform (most) GUC checks that need database access. Even when
only shared tables are needed, unless they're
nailed (c.f. RelationCacheInitializePhase2()), they cannot be accessed
without pg_class etc. being present.
Fix by extending the existing IsTransactionState() checks to also
check for MyDatabaseOid.
Reported-By: Vignesh C, Michael Paquier, Andres Freund
Author: Vignesh C, Andres Freund
Discussion: https://postgr.es/m/CALDaNm1KXK9gbZfY-p_peRFm_XrBh1OwQO1Kk6Gig0c0fVZ2uw%40mail.gmail.com
Backpatch: 9.4-
Alvaro Herrera [Mon, 10 Jun 2019 22:56:23 +0000 (18:56 -0400)]
Make pg_dump emit ATTACH PARTITION instead of PARTITION OF (reprise)
Using PARTITION OF can result in column ordering being changed from the
database being dumped, if the partition uses a column layout different
from the parent's. It's not pg_dump's job to editorialize on table
definitions, so this is not acceptable; back-patch all the way back to
pg10, where partitioned tables where introduced.
This change also ensures that partitions end up in the correct
tablespace, if different from the parent's; this is an oversight in ca4103025dfe (in pg12 only). Partitioned indexes (in pg11) don't have
this problem, because they're already created as independent indexes and
attached to their parents afterwards.
This change also has the advantage that the partition is restorable from
the dump (as a standalone table) even if its parent table isn't
restored.
The original commits (3b23552ad8bb in branch master) failed to cover
subsidiary column elements correctly, such as NOT NULL constraint and
CHECK constraints, as reported by Rushabh Lathia (initially as a failure
to restore serial columns). They were reverted. This recapitulation
commit fixes those problems.
Add some pg_dump tests to verify these things more exhaustively,
including constraints with legacy-inheritance tables, which were not
tested originally. In branches 10 and 11, add a local constraint to the
pg_dump test partition that was added by commit 2d7eeb1b1492 to master.
Fix operator naming in pg_trgm GUC option descriptions
Descriptions of pg_trgm GUC options have % replaced with %% like it was
a printf-like format. But that's not needed since they are just plain strings.
This commit fixed that. Backpatch to last supported version since this error
present from the beginning.
5871b884 introduced pg_trgm.word_similarity_threshold GUC, but its documentation
contains wrong indentation. This commit fixes that. Backpatch for easier
backpatching of other documentation fixes.
Discussion: https://postgr.es/m/4c735d30-ab59-fc0e-45d8-f90eb5ed3855%402ndquadrant.com
Author: Ian Barwick
Backpatch-through: 9.6
Fix copy-pasto in freeing memory on error in vacuumlo.
It's harmless to call PQfreemem() with a NULL argument, so the only
consequence was that if allocating 'schema' failed, but allocating 'table'
or 'field' succeeded, we would leak a bit of memory. That's highly
unlikely to happen, so this is just academical, but let's get it right.
Per bug #15838 from Timur Birsh. Backpatch back to 9.5, where the
PQfreemem() calls were introduced.
Amit Kapila [Thu, 6 Jun 2019 23:59:11 +0000 (05:29 +0530)]
Fix inconsistency in comments atop ExecParallelEstimate.
When this code was initially introduced in commit d1b7c1ff, the structure
used was SharedPlanStateInstrumentation, but later when it got changed to
Instrumentation structure in commit b287df70, we forgot to update the
comment.
Tom Lane [Mon, 3 Jun 2019 22:06:04 +0000 (18:06 -0400)]
Fix contrib/auto_explain to not cause problems in parallel workers.
A parallel worker process should not be making any decisions of its
own about whether to auto-explain. If the parent session process
passed down flags asking for instrumentation data, do that, otherwise
not. Trying to enable instrumentation anyway leads to bugs like the
"could not find key N in shm TOC" failure reported in bug #15821
from Christian Hofstaedtler.
We can implement this cheaply by piggybacking on the existing logic
for not doing anything when we've chosen not to sample a statement.
While at it, clean up some tin-eared coding related to the sampling
feature, including an off-by-one error that meant that asking for 1.0
sampling rate didn't actually result in sampling every statement.
Although the specific case reported here only manifested in >= v11,
I believe that related misbehaviors can be demonstrated in any version
that has parallel query; and the off-by-one error is certainly there
back to 9.6 where that feature was added. So back-patch to 9.6.
Tom Lane [Mon, 3 Jun 2019 20:59:16 +0000 (16:59 -0400)]
Fix unsafe memory management in CloneRowTriggersToPartition().
It's not really supported to call systable_getnext() in a different
memory context than systable_beginscan() was called in, and it's
*definitely* not safe to do so and then reset that context between
calls. I'm not very clear on how this code survived
CLOBBER_CACHE_ALWAYS testing ... but Alexander Lakhin found a case
that would crash it pretty reliably.
Per bug #15828. Fix, and backpatch to v11 where this code came in.
Michael Paquier [Sat, 1 Jun 2019 19:33:58 +0000 (15:33 -0400)]
Fix documentation of check_option in information_schema.views
Support of CHECK OPTION for updatable views has been added in 9.4, but
the documentation of information_schema never got the call even if the
information displayed is correct.
Tomas Vondra [Thu, 30 May 2019 14:16:12 +0000 (16:16 +0200)]
Make error logging in extended statistics more consistent
Most errors reported in extended statistics are internal issues, and so
should use elog(). The MCV list code was already following this rule, but
the functional dependencies and ndistinct coefficients were using a mix
of elog() and ereport(). Fix this by changing most places to elog(), with
the exception of input functions.
This is a mostly cosmetic change, it makes the life a little bit easier
for translators, as elog() messages are not translated. So backpatch to
PostgreSQL 10, where extended statistics were introduced.
Author: Tomas Vondra
Backpatch-through: 10 where extended statistics were added
Discussion: https://postgr.es/m/20190503154404.GA7478@alvherre.pgsql
Noah Misch [Tue, 28 May 2019 19:59:00 +0000 (12:59 -0700)]
In the pg_upgrade test suite, don't write to src/test/regress.
When this suite runs installcheck, redirect file creations from
src/test/regress to src/bin/pg_upgrade/tmp_check/regress. This closes a
race condition in "make -j check-world". If the pg_upgrade suite wrote
to a given src/test/regress/results file in parallel with the regular
src/test/regress invocation writing it, a test failed spuriously. Even
without parallelism, in "make -k check-world", the suite finishing
second overwrote the other's regression.diffs. This revealed test
"largeobject" assuming @abs_builddir@ is getcwd(), so fix that, too.
Buildfarm client REL_10, released fifty-four days ago, supports saving
regression.diffs from its new location. When an older client reports a
pg_upgradeCheck failure, it will no longer include regression.diffs.
Back-patch to 9.5, where pg_upgrade moved to src/bin.
Noah Misch [Tue, 28 May 2019 19:58:30 +0000 (12:58 -0700)]
In the pg_upgrade test suite, remove and recreate "tmp_check".
This allows "vcregress upgradecheck" to pass twice in immediate
succession, and it's more like how $(prove_check) works. Back-patch to
9.5, where pg_upgrade moved to src/bin.
Andres Freund [Thu, 23 May 2019 21:46:57 +0000 (14:46 -0700)]
pg_upgrade: Make test.sh's installcheck use to-be-upgraded version's bindir.
On master (after 700538) the old version's installed psql was used -
even when the old version might not actually be installed / might be
installed into a temporary directory. As commonly the case when just
executing make check for pg_upgrade, as $oldbindir is just the current
version's $bindir.
In the back branches, with --install specified, psql from the new
version's temporary installation was used, without --install (e.g for
NO_TEMP_INSTALL, cf 47b3c26642), the new version's installed psql was
used (which might or might not exist).
Author: Andres Freund
Discussion: https://postgr.es/m/20190522175150.c26f4jkqytahajdg@alap3.anarazel.de
Andrew Gierth [Thu, 23 May 2019 14:26:01 +0000 (15:26 +0100)]
Fix array size allocation for HashAggregate hash keys.
When there were duplicate columns in the hash key list, the array
sizes could be miscomputed, resulting in access off the end of the
array. Adjust the computation to ensure the array is always large
enough.
(I considered whether the duplicates could be removed in planning, but
I can't rule out the possibility that duplicate columns might have
different hash functions assigned. Simpler to just make sure it works
at execution time regardless.)
Bug apparently introduced in fc4b3dea2 as part of narrowing down the
tuples stored in the hashtable. Reported by Colm McHugh of Salesforce,
though I didn't use their patch. Backpatch back to version 10 where
the bug was introduced.
Michael Paquier [Thu, 23 May 2019 01:48:24 +0000 (10:48 +0900)]
Fix ordering of GRANT commands in pg_dumpall for tablespaces
This uses a method similar to 68a7c24f and now b8c6014 (applied for
database creation), which guarantees that GRANT commands using the WITH
GRANT OPTION are dumped in a way so as cascading dependencies are
respected. Note that tablespaces do not have support for initial
privileges via pg_init_privs, so the same method needs to be applied
again. It would be nice to merge all the logic generating ACL queries
in dumps under the same banner, but this requires extending the support
of pg_init_privs to objects that cannot use it yet, so this is left as
future work.
Discussion: https://postgr.es/m/20190522071555.GB1278@paquier.xyz
Author: Michael Paquier Reviewed-by: Nathan Bossart
Backpatch-through: 9.6
Michael Paquier [Wed, 22 May 2019 05:48:14 +0000 (14:48 +0900)]
Fix ordering of GRANT commands in pg_dump for database creation
This uses a method similar to 68a7c24f, which guarantees that GRANT
commands using the WITH GRANT OPTION are dumped in a way so as cascading
dependencies are respected. As databases do not have support for
initial privileges via pg_init_privs, we need to repeat again the same
ACL reordering method.
ACL for databases have been moved from pg_dumpall to pg_dump in v11, so
this impacts pg_dump for v11 and above, and pg_dumpall for v9.6 and
v10.
Andres Freund [Mon, 20 May 2019 01:01:06 +0000 (18:01 -0700)]
Minimally fix partial aggregation for aggregates that don't have one argument.
For partial aggregation combine steps,
AggStatePerTrans->numTransInputs was set to the transition function's
number of inputs, rather than the combine function's number of
inputs (always 1).
That lead to partial aggregates with strict combine functions to
wrongly check for NOT NULL input as required by strictness. When the
aggregate wasn't exactly passed one argument, the strictness check was
either omitted (in the 0 args case) or too many arguments were
checked. In the latter case we'd read beyond the end of
FunctionCallInfoData->args (only in master).
AggStatePerTrans->numTransInputs actually has been wrong since since
9.6, where partial aggregates were added. But it turns out to not be
an active problem in 9.6 and 10, because numTransInputs wasn't used at
all for combine functions: Before c253b722f6 there simply was no NULL
check for the input to strict trans functions, and after that the
check was simply hardcoded for the right offset in fcinfo, as it's
done by code specific to combine functions.
In bf6c614a2f2 (11) the strictness check was generalized, with common
code doing the strictness checks for both plain and combine transition
functions, based on numTransInputs. For combine functions this lead to
not emitting an expression step to check for strict input in the 0
arguments case, and in the > 1 arguments case, we'd check too many
arguments.Due to the fact that the relevant fcinfo->isnull[2..] was
always zero-initialized (more or less by accident, by being part of
the AggStatePerTrans struct, which is palloc0'ed), there was no
observable damage in the latter case before a9c35cf85ca1f, we just
checked too many array elements.
Due to the changes in a9c35cf85ca1f, > 1 argument bug became visible,
because these days fcinfo is a) dynamically allocated without being
zeroed b) exactly the length required for the number of specified
arguments (hardcoded to 2 in this case).
This commit only contains a fairly minimal fix, setting numTransInputs
to a hardcoded 1 when building a pertrans for a combine function. It
seems likely that we'll want to clean this up further (e.g. the
arguments build_pertrans_for_aggref() aren't particularly meaningful
for combine functions). But the wrap date for 12 beta1 is coming up
fast, so it seems good to have a minimal fix in place.
Backpatch to 11. While AggStatePerTrans->numTransInputs was set
wrongly before that, the value was not used for combine functions.
Reported-By: Rajkumar Raghuwanshi Diagnosed-By: Kyotaro Horiguchi, Jeevan Chalke, Andres Freund, David Rowley
Author: David Rowley, Kyotaro Horiguchi, Andres Freund
Discussion: https://postgr.es/m/CAKcux6=uZEyWyLw0N7HtR9OBc-sWEFeByEZC7t-KDf15FKxVew@mail.gmail.com
Michael Paquier [Mon, 20 May 2019 00:48:27 +0000 (09:48 +0900)]
Fix some grammar in documentation of spgist and pgbench
Discussion: https://postgr.es/m/92961161-9b49-e42f-0a72-d5d47e0ed4de@postgrespro.ru
Author: Liudmila Mantrova Reviewed-by: Jonathan Katz, Tom Lane, Michael Paquier
Backpatch-through: 9.4
Noah Misch [Sun, 19 May 2019 21:36:44 +0000 (14:36 -0700)]
In the pg_upgrade test suite, don't write to src/test/regress.
When this suite runs installcheck, redirect file creations from
src/test/regress to src/bin/pg_upgrade/tmp_check/regress. This closes a
race condition in "make -j check-world". If the pg_upgrade suite wrote
to a given src/test/regress/results file in parallel with the regular
src/test/regress invocation writing it, a test failed spuriously. Even
without parallelism, in "make -k check-world", the suite finishing
second overwrote the other's regression.diffs. This revealed test
"largeobject" assuming @abs_builddir@ is getcwd(), so fix that, too.
Buildfarm client REL_10, released forty-five days ago, supports saving
regression.diffs from its new location. When an older client reports a
pg_upgradeCheck failure, it will no longer include regression.diffs.
Back-patch to 9.5, where pg_upgrade moved to src/bin.
Tom Lane [Fri, 17 May 2019 23:44:19 +0000 (19:44 -0400)]
Restructure creation of run-time pruning steps.
Previously, gen_partprune_steps() always built executor pruning steps
using all suitable clauses, including those containing PARAM_EXEC
Params. This meant that the pruning steps were only completely safe
for executor run-time (scan start) pruning. To prune at executor
startup, we had to ignore the steps involving exec Params. But this
doesn't really work in general, since there may be logic changes
needed as well --- for example, pruning according to the last operator's
btree strategy is the wrong thing if we're not applying that operator.
The rules embodied in gen_partprune_steps() and its minions are
sufficiently complicated that tracking their incremental effects in
other logic seems quite impractical.
Short of a complete redesign, the only safe fix seems to be to run
gen_partprune_steps() twice, once to create executor startup pruning
steps and then again for run-time pruning steps. We can save a few
cycles however by noting during the first scan whether we rejected
any clauses because they involved exec Params --- if not, we don't
need to do the second scan.
In support of this, refactor the internal APIs in partprune.c to make
more use of passing information in the GeneratePruningStepsContext
struct, rather than as separate arguments.
This is, I hope, the last piece of our response to a bug report from
Alan Jackson. Back-patch to v11 where this code came in.
Tom Lane [Thu, 16 May 2019 15:58:22 +0000 (11:58 -0400)]
Fix partition pruning to treat stable comparison operators properly.
Cross-type comparison operators in a btree or hash opclass might be
only stable not immutable (this is true of timestamp vs. timestamptz
for example). partprune.c ignored this possibility and would perform
plan-time pruning with them anyway, possibly leading to wrong answers
if the environment changed between planning and execution.
To fix, teach gen_partprune_steps() to do things differently when
creating plan-time pruning steps vs. run-time pruning steps.
analyze_partkey_exprs() also needs an extra check, which is rather
annoying but now is not the time to restructure things enough to
avoid that.
While at it, simplify the logic for the plan-time case a little
by insisting that the comparison value be a Const and nothing else.
This relies on the assumption that eval_const_expressions will have
reduced any immutable expression to a Const; which is not quite
100% true, but certainly any case that comes up often enough to be
interesting should have simplification logic there.
Also improve a bunch of inadequate/obsolete/wrong comments.
Per discussion of a report from Alan Jackson (though this fixes only one
aspect of that problem). Back-patch to v11 where this code came in.
Andres Freund [Tue, 14 May 2019 18:45:40 +0000 (11:45 -0700)]
Add isolation test for INSERT ON CONFLICT speculative insertion failure.
This path previously was not reliably covered. There was some
heuristic coverage via insert-conflict-toast.spec, but that test is
not deterministic, and only tested for a somewhat specific bug.
Backpatch, as this is a complicated and otherwise untested code
path. Unfortunately 9.5 cannot handle two waiting sessions, and thus
cannot execute this test.
Triggered by a conversion with Melanie Plageman.
Author: Andres Freund
Discussion: https://postgr.es/m/CAAKRu_a7hbyrk=wveHYhr4LbcRnRCG=yPUVoQYB9YO1CdUBE9Q@mail.gmail.com
Backpatch: 9.5-
Peter Geoghegan [Mon, 13 May 2019 22:39:05 +0000 (15:39 -0700)]
Doc: Refer to line pointers as item identifiers.
An upcoming HEAD-only patch will standardize the terminology around
ItemIdData variables/line pointers, ending the practice of referring to
them as "item pointers". Make the "Database Page Layout" docs
consistent with the new policy. The term "item identifier" is already
used in the same section, so stick with that.
Discussion: https://postgr.es/m/CAH2-Wz=c=MZQjUzde3o9+2PLAPuHTpVZPPdYxN=E4ndQ2--8ew@mail.gmail.com
Backpatch: All supported branches.
Tom Lane [Mon, 13 May 2019 21:23:00 +0000 (17:23 -0400)]
Fix logical replication's ideas about which type OIDs are built-in.
Only hand-assigned type OIDs should be presumed to match across different
PG servers; those assigned during genbki.pl or during initdb are likely
to change due to addition or removal of unrelated objects.
This means that the cutoff should be FirstGenbkiObjectId (in HEAD)
or FirstBootstrapObjectId (before that), not FirstNormalObjectId.
Compare postgres_fdw's is_builtin() test.
It's likely that this error has no observable consequence in a
normally-functioning system, since ATM the only affected type OIDs are
system catalog rowtypes and information_schema types, which would not
typically be interesting for logical replication. But you could
probably break it if you tried hard, so back-patch.
Peter Geoghegan [Mon, 13 May 2019 17:27:57 +0000 (10:27 -0700)]
Don't leave behind junk nbtree pages during split.
Commit 8fa30f906be reduced the elevel of a number of "can't happen"
_bt_split() errors from PANIC to ERROR. At the same time, the new right
page buffer for the split could continue to be acquired well before the
critical section. This was possible because it was relatively
straightforward to make sure that _bt_split() could not throw an error,
with a few specific exceptions. The exceptional cases were safe because
they involved specific, well understood errors, making it possible to
consistently zero the right page before actually raising an error using
elog(). There was no danger of leaving around a junk page, provided
_bt_split() stuck to this coding rule.
Commit 8224de4f, which introduced INCLUDE indexes, added code to make
_bt_split() truncate away non-key attributes. This happened at a point
that broke the rule around zeroing the right page in _bt_split(). If
truncation failed (perhaps due to palloc() failure), that would result
in an errant right page buffer with junk contents. This could confuse
VACUUM when it attempted to delete the page, and should be avoided on
general principle.
To fix, reorganize _bt_split() so that truncation occurs before the new
right page buffer is even acquired. A junk page/buffer will not be left
behind if _bt_nonkey_truncate()/_bt_truncate() raise an error.
Discussion: https://postgr.es/m/CAH2-WzkcWT_-NH7EeL=Az4efg0KCV+wArygW8zKB=+HoP=VWMw@mail.gmail.com
Backpatch: 11-, where INCLUDE indexes were introduced.
Tom Lane [Mon, 13 May 2019 14:53:19 +0000 (10:53 -0400)]
Fix misuse of an integer as a bool.
pgtls_read_pending is declared to return bool, but what the underlying
SSL_pending function returns is a count of available bytes.
This is actually somewhat harmless if we're using C99 bools, but in
the back branches it's a live bug: if the available-bytes count happened
to be a multiple of 256, it would get converted to a zero char value.
On machines where char is signed, counts of 128 and up could misbehave
as well. The net effect is that when using SSL, libpq might block
waiting for data even though some has already been received.
Broken by careless refactoring in commit 4e86f1b16, so back-patch
to 9.5 where that came in.
Tom Lane [Sun, 12 May 2019 22:53:12 +0000 (18:53 -0400)]
Fix misoptimization of "{1,1}" quantifiers in regular expressions.
A bounded quantifier with m = n = 1 might be thought a no-op. But
according to our documentation (which traces back to Henry Spencer's
original man page) it still imposes greediness, or non-greediness in the
case of the non-greedy variant "{1,1}?", on whatever it's attached to.
This turns out not to work though, because parseqatom() optimizes away
the m = n = 1 case without regard for whether it's supposed to change
the greediness of the argument RE.
We can fix this by just not applying the optimization when the greediness
needs to change; the subsequent general cases handle it fine.
The three cases in which we can still apply the optimization are
(a) no quantifier, or quantifier does not impose a preference;
(b) atom has no greediness property, implying it cannot match a
variable amount of text anyway; or
(c) quantifier's greediness is same as atom's.
Note that in most cases where one of these applies, we'd have exited
earlier in the "not a messy case" fast path. I think it's now only
possible to get to the optimization when the atom involves capturing
parentheses or a non-top-level backref.
Back-patch to all supported branches. I'd ordinarily be hesitant to
put a subtle behavioral change into back branches, but in this case
it's very hard to see a reason why somebody would write "{1,1}?" unless
they're trying to get the documented change-of-greediness behavior.
Noah Misch [Sun, 12 May 2019 17:33:05 +0000 (10:33 -0700)]
Fail pgwin32_message_to_UTF16() for SQL_ASCII messages.
The function had been interpreting SQL_ASCII messages as UTF8, throwing
an error when they were invalid UTF8. The new behavior is consistent
with pg_do_encoding_conversion(). This affects LOG_DESTINATION_STDERR
and LOG_DESTINATION_EVENTLOG, which will send untranslated bytes to
write() and ReportEventA(). On buildfarm member bowerbird, enabling
log_connections caused an error whenever the role name was not valid
UTF8. Back-patch to 9.4 (all supported versions).
Tom Lane [Sun, 12 May 2019 01:27:13 +0000 (21:27 -0400)]
Rearrange pgstat_bestart() to avoid failures within its critical section.
We long ago decided to design the shared PgBackendStatus data structure to
minimize the cost of writing status updates, which means that writers just
have to increment the st_changecount field twice. That isn't hooked into
any sort of resource management mechanism, which means that if something
were to throw error between the two increments, the st_changecount field
would be left odd indefinitely. That would cause readers to lock up.
Now, since it's also a bad idea to leave the field odd for longer than
absolutely necessary (because readers will spin while we have it set),
the expectation was that we'd treat these segments like spinlock critical
sections, with only short, more or less straight-line, code in them.
That was fine as originally designed, but commit 9029f4b37 broke it
by inserting a significant amount of non-straight-line code into
pgstat_bestart(), code that is very capable of throwing errors, not to
mention taking a significant amount of time during which readers will spin.
We have a report from Neeraj Kumar of readers actually locking up, which
I suspect was due to an encoding conversion error in X509_NAME_to_cstring,
though conceivably it was just a garden-variety OOM failure.
Subsequent commits have loaded even more dubious code into pgstat_bestart's
critical section (and commit fc70a4b0d deserves some kind of booby prize
for managing to miss the critical section entirely, although the negative
consequences seem minimal given that the PgBackendStatus entry should be
seen by readers as inactive at that point).
The right way to fix this mess seems to be to compute all these values
into a local copy of the process' PgBackendStatus struct, and then just
copy the data back within the critical section proper. This plan can't
be implemented completely cleanly because of the struct's heavy reliance
on out-of-line strings, which we must initialize separately within the
critical section. But still, the critical section is far smaller and
safer than it was before.
In hopes of forestalling future errors of the same ilk, rename the
macros for st_changecount management to make it more apparent that
the writer-side macros create a critical section. And to prevent
the worst consequences if we nonetheless manage to mess it up anyway,
adjust those macros so that they really are a critical section, ie
they now bump CritSectionCount. That doesn't add much overhead, and
it guarantees that if we do somehow throw an error while the counter
is odd, it will lead to PANIC and a database restart to reset shared
memory.
Back-patch to 9.5 where the problem was introduced.
In HEAD, also fix an oversight in commit b0b39f72b: it failed to teach
pgstat_read_current_status to copy st_gssstatus data from shared memory to
local memory. Hence, subsequent use of that data within the transaction
would potentially see changing data that it shouldn't see.
Noah Misch [Sat, 11 May 2019 07:22:38 +0000 (00:22 -0700)]
Honor TEMP_CONFIG in TAP suites.
The buildfarm client uses TEMP_CONFIG to implement its extra_config
setting. Except for stats_temp_directory, extra_config now applies to
TAP suites; extra_config values seen in the past month are compatible
with this. Back-patch to 9.6, where PostgresNode was introduced, so the
buildfarm can rely on it sooner.
Tom Lane [Fri, 10 May 2019 18:56:41 +0000 (14:56 -0400)]
Cope with EINVAL and EIDRM shmat() failures in PGSharedMemoryAttach.
There's a very old race condition in our code to see whether a pre-existing
shared memory segment is still in use by a conflicting postmaster: it's
possible for the other postmaster to remove the segment in between our
shmctl() and shmat() calls. It's a narrow window, and there's no risk
unless both postmasters are using the same port number, but that's possible
during parallelized "make check" tests. (Note that while the TAP tests
take some pains to choose a randomized port number, pg_regress doesn't.)
If it does happen, we treated that as an unexpected case and errored out.
To fix, allow EINVAL to be treated as segment-not-present, and the same
for EIDRM on Linux. AFAICS, the considerations here are basically
identical to the checks for acceptable shmctl() failures, so I documented
and coded it that way.
While at it, adjust PGSharedMemoryAttach's API to remove its undocumented
dependency on UsedShmemSegAddr in favor of passing the attach address
explicitly. This makes it easier to be sure we're using a null shmaddr
when probing for segment conflicts (thus avoiding questions about what
EINVAL means). I don't think there was a bug there, but it required
fragile assumptions about the state of UsedShmemSegAddr during
PGSharedMemoryIsInUse.
Commit c09850992 may have made this failure more probable by applying
the conflicting-segment tests more often. Hence, back-patch to all
supported branches, as that was.
Tom Lane [Thu, 9 May 2019 20:52:49 +0000 (16:52 -0400)]
Repair issues with faulty generation of merge-append plans.
create_merge_append_plan failed to honor the CP_EXACT_TLIST flag:
it would generate the expected targetlist but then it felt free to
add resjunk sort targets to it. This demonstrably leads to assertion
failures in v11 and HEAD, and it's probably just accidental that we
don't see the same in older branches. I've not looked into whether
there would be any real-world consequences in non-assert builds.
In HEAD, create_append_plan has sprouted the same problem, so fix
that too (although we do not have any test cases that seem able to
reach that bug). This is an oversight in commit 3fc6e2d7f which
invented the CP_EXACT_TLIST flag, so back-patch to 9.6 where that
came in.
convert_subquery_pathkeys would create pathkeys for subquery output
values if they match any EquivalenceClass known in the outer query
and are available in the subquery's syntactic targetlist. However,
the second part of that condition is wrong, because such values might
not appear in the subquery relation's reltarget list, which would
mean that they couldn't be accessed above the level of the subquery
scan. We must check that they appear in the reltarget list, instead.
This can lead to dropping knowledge about the subquery's sort
ordering, but I believe it's okay, because any sort key that the
outer query actually has any interest in would appear in the
reltarget list.
This second issue is of very long standing, but right now there's no
evidence that it causes observable problems before 9.6, so I refrained
from back-patching further than that. We can revisit that choice if
somebody finds a way to make it cause problems in older branches.
(Developing useful test cases for these issues is really problematic;
fixing convert_subquery_pathkeys removes the only known way to exhibit
the create_merge_append_plan bug, and neither of the test cases added
by this patch causes a problem in all branches, even when considering
the issues separately.)
The second issue explains bug #15795 from Suresh Kumar R ("could not
find pathkey item to sort" with nested DISTINCT queries). I stumbled
across the first issue while investigating that.
Michael Paquier [Thu, 9 May 2019 01:29:29 +0000 (10:29 +0900)]
Fix error status of vacuumdb when multiple jobs are used
When running a batch of VACUUM or ANALYZE commands on a given database,
there were cases where it is possible to have vacuumdb not report an
error where it actually should, leading to incorrect status results.
Author: Julien Rouhaud Reviewed-by: Amit Kapila, Michael Paquier
Discussion: https://postgr.es/m/CAOBaU_ZuTwz7CtqLYJ1Ouuh272bTQPLN8b1bAPk0bCBm4PDMTQ@mail.gmail.com
Backpatch-through: 9.5
Fujii Masao [Wed, 8 May 2019 16:35:13 +0000 (01:35 +0900)]
Fix documentation for the privileges required for replication functions.
Previously it's documented that use of replication functions is
restricted to superusers. This is true for the functions which
use replication origin, but not for pg_logicl_emit_message() and
functions which use replication slot. For example, not only
superusers but also users with REPLICATION privilege is allowed
to use the functions for replication slot. This commit fixes
the documentation for the privileges required for those replication
functions.
Thomas Munro [Mon, 6 May 2019 03:02:41 +0000 (15:02 +1200)]
Probe only 127.0.0.1 when looking for ports on Unix.
Commit c0985099, later adjusted by commit 4ab02e81, probed 0.0.0.0
in addition to 127.0.0.1, for the benefit of Windows build farm
animals. It isn't really useful on Unix systems, and turned out to
be a bit inconvenient to users of some corporate firewall software.
Switch back to probing just 127.0.0.1 on non-Windows systems.
Michael Paquier [Tue, 7 May 2019 05:19:56 +0000 (14:19 +0900)]
Remove some code related to 7.3 and older servers from tools of src/bin/
This code was broken as of 582edc3, and is most likely not used anymore.
Note that pg_dump supports servers down to 8.0, and psql has code to
support servers down to 7.4.
Author: Julien Rouhaud Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CAOBaU_Y5y=zo3+2gf+2NJC1pvMYPcbRXoQaPXx=U7+C8Qh4CzQ@mail.gmail.com
Alvaro Herrera [Mon, 6 May 2019 16:23:49 +0000 (12:23 -0400)]
Revert "Make pg_dump emit ATTACH PARTITION instead of PARTITION OF"
... and fallout (from branches 10, 11 and master). The change was
ill-considered, and it broke a few normal use cases; since we don't have
time to fix it, we'll try again after this week's minor releases.
Michael Paquier [Mon, 6 May 2019 12:44:39 +0000 (21:44 +0900)]
Fix tuple printing in error message of tuple routing for partitions
With correctly crafted DDLs, this could lead to disclosure of arbitrary
backend memory a user may have no right to access. This impacts only
REL_11_STABLE, as the issue has been introduced by 34295b8.
On HEAD, add regression tests to cover this issue in the future.
Author: Michael Paquier Reviewed-by: Noah Misch
Security: CVE-2019-10129
Dean Rasheed [Mon, 6 May 2019 10:56:37 +0000 (11:56 +0100)]
Use checkAsUser for selectivity estimator checks, if it's set.
In examine_variable() and examine_simple_variable(), when checking the
user's table and column privileges to determine whether to grant
access to the pg_statistic data, use checkAsUser for the privilege
checks, if it's set. This will be the case if we're accessing the
table via a view, to indicate that we should perform privilege checks
as the view owner rather than the current user.
This change makes this planner check consistent with the check in the
executor, so the planner will be able to make use of statistics if the
table is accessible via the view. This fixes a performance regression
introduced by commit e2d4ef8de8, which affects queries against
non-security barrier views in the case where the user doesn't have
privileges on the underlying table, but the view owner does.
Note that it continues to provide the same safeguards controlling
access to pg_statistic for direct table access (in which case
checkAsUser won't be set) and for security barrier views, because of
the nearby checks on rte->security_barrier and rte->securityQuals.
Back-patch to all supported branches because e2d4ef8de8 was.
Dean Rasheed, reviewed by Jonathan Katz and Stephen Frost.
Dean Rasheed [Mon, 6 May 2019 10:41:22 +0000 (11:41 +0100)]
Fix security checks for selectivity estimation functions with RLS.
In commit e2d4ef8de8, security checks were added to prevent
user-supplied operators from running over data from pg_statistic
unless the user has table or column privileges on the table, or the
operator is leakproof. For a table with RLS, however, checking for
table or column privileges is insufficient, since that does not
guarantee that the user has permission to view all of the column's
data.
Fix this by also checking for securityQuals on the RTE, and insisting
that the operator be leakproof if there are any. Thus the
leakproofness check will only be skipped if there are no securityQuals
and the user has table or column privileges on the table -- i.e., only
if we know that the user has access to all the data in the column.
Back-patch to 9.5 where RLS was added.
Dean Rasheed, reviewed by Jonathan Katz and Stephen Frost.
Andres Freund [Mon, 6 May 2019 06:31:58 +0000 (23:31 -0700)]
Remove reindex_catalog test from test schedules.
As the test currently causes occasional deadlocks (due to the schema
cleanup from previous sessions potentially still running), and the
patch from f912d7dec2 has gotten a fair bit of buildfarm coverage,
remove the test from the test schedules. There's a set of minor
releases coming up.
Leave the tests in place, so it can manually be run using EXTRA_TESTS.
For now also leave it in master, as there's no imminent release, and
there's plenty (re-)index related work in 12. But we'll have to
disable it before long there too, unless somebody comes up with simple
enough fixes for the deadlock (I'm about to post a vague idea to the
list).
Discussion: https://postgr.es/m/4622.1556982247@sss.pgh.pa.us
Backpatch: 9.4-11 (no master!)
Tom Lane [Sun, 5 May 2019 17:10:07 +0000 (13:10 -0400)]
Fix style violations in syscache lookups.
Project style is to check the success of SearchSysCacheN and friends
by applying HeapTupleIsValid to the result. A tiny minority of calls
creatively did it differently. Bring them into line with the rest.
This is just cosmetic, since HeapTupleIsValid is indeed just a null
check at the moment ... but that may not be true forever, and in any
case it puts a mental burden on readers who may wonder why these
call sites are not like the rest.
Back-patch to v11 just to keep the branches in sync. (The bulk of these
errors seem to have originated in v11 or v12, though a few are old.)
Per searching to see if anyplace else had made the same error
repaired in 62148c352.
Tom Lane [Sun, 5 May 2019 16:44:32 +0000 (12:44 -0400)]
Add check for syscache lookup failure in update_relispartition().
Omitted in commit 05b38c7e6 (though it looks like the original blame
belongs to 9e9befac4). A failure is admittedly unlikely, but if it
did happen, SIGSEGV is not the approved method of reporting it.
Per Coverity. Back-patch to v11 where the broken code originated.
Tom Lane [Thu, 2 May 2019 23:11:28 +0000 (19:11 -0400)]
Fix reindexing of pg_class indexes some more.
Commits 3dbb317d3 et al failed under CLOBBER_CACHE_ALWAYS testing.
Investigation showed that to reindex pg_class_oid_index, we must
suppress accesses to the index (via SetReindexProcessing) before we call
RelationSetNewRelfilenode, or at least before we do CommandCounterIncrement
therein; otherwise, relcache reloads happening within the CCI may try to
fetch pg_class rows using the index's new relfilenode value, which is as
yet an empty file.
Of course, the point of 3dbb317d3 was that that ordering didn't work
either, because then RelationSetNewRelfilenode's own update of the index's
pg_class row cannot access the index, should it need to.
There are various ways we might have got around that, but Andres Freund
came up with a brilliant solution: for a mapped index, we can really just
skip the pg_class update altogether. The only fields it was actually
changing were relpages etc, but it was just setting them to zeroes which
is useless make-work. (Correct new values will be installed at the end
of index build.) All pg_class indexes are mapped and probably always will
be, so this eliminates the problem by removing work rather than adding it,
always a pleasant outcome. Having taught RelationSetNewRelfilenode to do
it that way, we can revert the code reordering in reindex_index. (But
I left the moved setup code where it was; there seems no reason why it
has to run without use of the old index. If you're trying to fix a
busted pg_class index, you'll have had to disable system index use
altogether to get this far.)
Moreover, this means we don't need RelationSetIndexList at all, because
reindex_relation's hacking to make "REINDEX TABLE pg_class" work is
likewise now unnecessary. We'll leave that code in place in the back
branches, but a follow-on patch will remove it in HEAD.
In passing, do some minor cleanup for commit 5c1560606 (in HEAD only),
notably removing a duplicate newrnode assignment.
Patch by me, using a core idea due to Andres Freund. Back-patch to all
supported branches, as 3dbb317d3 was.
Andres Freund [Wed, 1 May 2019 00:45:32 +0000 (17:45 -0700)]
Run catalog reindexing test from 3dbb317d32 serially, to avoid deadlocks.
The tests turn out to cause deadlocks in some circumstances. Fairly
reproducibly so with -DRELCACHE_FORCE_RELEASE
-DCATCACHE_FORCE_RELEASE. Some of the deadlocks may be hard to fix
without disproportionate measures, but others probably should be fixed
- but not in 12.
We discussed removing the new tests until we can fix the issues
underlying the deadlocks, but results from buildfarm animal
markhor (which runs with CLOBBER_CACHE_ALWAYS) indicates that there
might be a more severe, as of yet undiagnosed, issue (including on
stable branches) with reindexing catalogs. The failure is:
ERROR: could not read block 0 in file "base/16384/28025": read only 0 of 8192 bytes
Therefore it seems advisable to keep the tests.
It's not certain that running the tests in isolation removes the risk
of deadlocks. It's possible that additional locks are needed to
protect against a concurrent auto-analyze or such.
Per discussion with Tom Lane.
Discussion: https://postgr.es/m/28926.1556664156@sss.pgh.pa.us
Backpatch: 9.4-, like 3dbb317d3
Tom Lane [Tue, 30 Apr 2019 19:03:35 +0000 (15:03 -0400)]
Clean up handling of constraint_exclusion and enable_partition_pruning.
The interaction of these parameters was a bit confused/confusing,
and in fact v11 entirely misses the opportunity to apply partition
constraints when a partition is accessed directly (rather than
indirectly from its parent).
In HEAD, establish the principle that enable_partition_pruning controls
partition pruning and nothing else. When accessing a partition via its
parent, we do partition pruning (if enabled by enable_partition_pruning)
and then there is no need to consider partition constraints in the
constraint_exclusion logic. When accessing a partition directly, its
partition constraints are applied by the constraint_exclusion logic,
only if constraint_exclusion = on.
In v11, we can't have such a clean division of these GUCs' effects,
partly because we don't want to break compatibility too much in a
released branch, and partly because the clean coding requires
inheritance_planner to have applied partition pruning to a partitioned
target table, which it doesn't in v11. However, we can tweak things
enough to cover the missed case, which seems like a good idea since
it's potentially a performance regression from v10. This patch keeps
v11's previous behavior in which enable_partition_pruning overrides
constraint_exclusion for an inherited target table, though.
In HEAD, also teach relation_excluded_by_constraints that it's okay to use
inheritable constraints when trying to prune a traditional inheritance
tree. This might not be thought worthy of effort given that that feature
is semi-deprecated now, but we have enough infrastructure that it only
takes a couple more lines of code to do it correctly.
Andres Freund [Tue, 30 Apr 2019 02:39:36 +0000 (19:39 -0700)]
Fix potential assertion failure when reindexing a pg_class index.
When reindexing individual indexes on pg_class it was possible to
either trigger an assertion failure:
TRAP: FailedAssertion("!(!ReindexIsProcessingIndex(((index)->rd_id)))
That's because reindex_index() called SetReindexProcessing() - which
enables an asserts ensuring no index insertions happen into the index
- before calling RelationSetNewRelfilenode(). That not correct for
indexes on pg_class, because RelationSetNewRelfilenode() updates the
relevant pg_class row, which needs to update the indexes.
The are two reasons this wasn't noticed earlier. Firstly the bug
doesn't trigger when reindexing all of pg_class, as reindex_relation
has code "hiding" all yet-to-be-reindexed indexes. Secondly, the bug
only triggers when the the update to pg_class doesn't turn out to be a
HOT update - otherwise there's no index insertion to trigger the
bug. Most of the time there's enough space, making this bug hard to
trigger.
To fix, move RelationSetNewRelfilenode() to before the
SetReindexProcessing() (and, together with some other code, to outside
of the PG_TRY()).
To make sure the error checking intended by SetReindexProcessing() is
more robust, modify CatalogIndexInsert() to check
ReindexIsProcessingIndex() even when the update is a HOT update.
Also add a few regression tests for REINDEXing of system catalogs.
The last two improvements would have prevented some of the issues
fixed in 5c1560606dc4c from being introduced in the first place.
Reported-By: Michael Paquier Diagnosed-By: Tom Lane and Andres Freund
Author: Andres Freund Reviewed-By: Tom Lane
Discussion: https://postgr.es/m/20190418011430.GA19133@paquier.xyz
Backpatch: 9.4-, the bug is present in all branches
Peter Eisentraut [Mon, 29 Apr 2019 06:44:51 +0000 (08:44 +0200)]
Fix potential catalog corruption with temporary identity columns
If a temporary table with an identity column and ON COMMIT DROP is
created in a single-statement transaction (not useful, but allowed),
it would leave the catalog corrupted. We need to add a
CommandCounterIncrement() so that PreCommit_on_commit_actions() sees
the created dependency between table and sequence and can clean it
up.
The analogous and more useful case of doing this in a transaction
block already runs some CommandCounterIncrement() before it gets to
the on-commit cleanup, so it wasn't a problem in practical use.
Several locations for placing the new CommandCounterIncrement() call
were discussed. This patch places it at the end of
standard_ProcessUtility(). That would also help if other commands
were to create catalog entries that some on-commit action would like
to see.
Bug: #15631 Reported-by: Serge Latyntsev <dnsl48@gmail.com>
Author: Peter Eisentraut <peter.eisentraut@2ndquadrant.com> Reviewed-by: Michael Paquier <michael@paquier.xyz>
Tom Lane [Sat, 27 Apr 2019 17:15:54 +0000 (13:15 -0400)]
Avoid postgres_fdw crash for a targetlist entry that's just a Param.
foreign_grouping_ok() is willing to put fairly arbitrary expressions into
the targetlist of a remote SELECT that's doing grouping or aggregation on
the remote side, including expressions that have no foreign component to
them at all. This is possibly a bit dubious from an efficiency standpoint;
but it rises to the level of a crash-causing bug if the expression is just
a Param or non-foreign Var. In that case, the expression will necessarily
also appear in the fdw_exprs list of values we need to send to the remote
server, and then setrefs.c's set_foreignscan_references will mistakenly
replace the fdw_exprs entry with a Var referencing the targetlist result.
The root cause of this problem is bad design in commit e7cb7ee14: it put
logic into set_foreignscan_references that IMV is postgres_fdw-specific,
and yet this bug shows that it isn't postgres_fdw-specific enough. The
transformation being done on fdw_exprs assumes that fdw_exprs is to be
evaluated with the fdw_scan_tlist as input, which is not how postgres_fdw
uses it; yet it could be the right thing for some other FDW. (In the
bigger picture, setrefs.c has no business assuming this for the other
expression fields of a ForeignScan either.)
The right fix therefore would be to expand the FDW API so that the
FDW could inform setrefs.c how it intends to evaluate these various
expressions. We can't change that in the back branches though, and we
also can't just summarily change setrefs.c's behavior there, or we're
likely to break external FDWs.
As a stopgap, therefore, hack up postgres_fdw so that it won't attempt
to send targetlist entries that look exactly like the fdw_exprs entries
they'd produce. In most cases this actually produces a superior plan,
IMO, with less data needing to be transmitted and returned; so we probably
ought to think harder about whether we should ship tlist expressions at
all when they don't contain any foreign Vars or Aggs. But that's an
optimization not a bug fix so I left it for later. One case where this
produces an inferior plan is where the expression in question is actually
a GROUP BY expression: then the restriction prevents us from using remote
grouping. It might be possible to work around that (since that would
reduce to group-by-a-constant on the remote side); but it seems like a
pretty unlikely corner case, so I'm not sure it's worth expending effort
solely to improve that. In any case the right long-term answer is to fix
the API as sketched above, and then revert this hack.
Per bug #15781 from Sean Johnston. Back-patch to v10 where the problem
was introduced.
Tom Lane [Fri, 26 Apr 2019 23:46:26 +0000 (19:46 -0400)]
Sync our copy of the timezone library with IANA release tzcode2019a.
This corrects a small bug in zic that caused it to output an incorrect
year-2440 transition in the Africa/Casablanca zone.
More interestingly, zic has grown a "-r" option that limits the range of
zone transitions that it will put into the output files. That might be
useful to people who don't like the weird GMT offsets that tzdb likes
to use for very old dates. It appears that for dates before the cutoff
time specified with -r, zic will use the zone's standard-time offset
as of the cutoff time. So for example one might do
to cause all dates before 1910-01-01 to be treated as though 1910
standard time prevailed indefinitely far back. (Don't blame me for
the unfriendly way of specifying the cutoff time --- it's seconds
since or before the Unix epoch. You can use extract(epoch ...)
to calculate it.)
Tom Lane [Fri, 26 Apr 2019 21:56:26 +0000 (17:56 -0400)]
Update time zone data files to tzdata release 2019a.
DST law changes in Palestine and Metlakatla.
Historical corrections for Israel.
Etc/UCT is now a backward-compatibility link to Etc/UTC, instead
of being a separate zone that generates the abbreviation "UCT",
which nowadays is typically a typo. Postgres will still accept
"UCT" as an input zone name, but it won't output it.
Tom Lane [Fri, 26 Apr 2019 21:18:07 +0000 (17:18 -0400)]
Apply stopgap fix for bug #15672.
Fix DefineIndex so that it doesn't attempt to pass down a to-be-reused
index relfilenode to a child index creation, and fix TryReuseIndex
to not think that reuse is sensible for a partitioned index.
In v11, this fixes a problem where ALTER TABLE on a partitioned table
could assign the same relfilenode to several different child indexes,
causing very nasty catalog corruption --- in fact, attempting to DROP
the partitioned table then leads not only to a database crash, but to
inability to restart because the same crash will recur during WAL replay.
Either of these two changes would be enough to prevent the failure, but
since neither action could possibly be sane, let's put in both changes
for future-proofing.
In HEAD, no such bug manifests, but that's just an accidental consequence
of having changed the pg_class representation of partitioned indexes to
have relfilenode = 0. Both of these changes still seem like smart
future-proofing.
This is only a stop-gap because the code for ALTER TABLE on a partitioned
table with a no-op type change still leaves a great deal to be desired.
As the added regression tests show, it gets things wrong for comments on
child indexes/constraints, and it is regenerating child indexes it doesn't
have to. However, fixing those problems will take more work which may not
get back-patched into v11. We need a fix for the corruption problem now.
Per bug #15672 from Jianing Yang.
Patch by me, regression test cases based on work by Amit Langote,
who also did a lot of the investigative work.
Add FDW documentation notes about insert and update tuple routing and COPY.
Author: Laurenz Albe and Etsuro Fujita Reviewed-by: Laurenz Albe and Amit Langote
Backpatch-through: 11 where support for that by FDWs was added
Discussion: https://postgr.es/m/bf36a0288e8f31b4f2f40952e225bf892dc1ffc5.camel@cybertec.at
When an existing index in a partition is attached to a new index on
its parent, we forgot to set the "relispartition" flag correctly, which
meant that it was not possible to find the index in various operations,
such as adding a foreign key constraint that references that partitioned
table. One of four places that was assigning the parent index was
forgetting to do that, so fix by shifting responsibility of updating the
flag to the routine that changes the parent.
Make pg_dump emit ATTACH PARTITION instead of PARTITION OF
Using PARTITION OF can result in column ordering being changed from the
database being dumped, if the partition uses a column layout different
from the parent's. It's not pg_dump's job to editorialize on table
definitions, so this is not acceptable; back-patch all the way back to
pg10, where partitioned tables where introduced.
This change also ensures that partitions end up in the correct
tablespace, if different from the parent's; this is an oversight in ca4103025dfe (in pg12 only). Partitioned indexes (in pg11) don't have
this problem, because they're already created as independent indexes and
attached to their parents afterwards.
This change also has the advantage that the partition is restorable from
the dump (as a standalone table) even if its parent table isn't
restored.
Author: David Rowley Reviewed-by: Álvaro Herrera
Discussion: https://postgr.es/m/CAKJS1f_1c260nOt_vBJ067AZ3JXptXVRohDVMLEBmudX1YEx-A@mail.gmail.com
Discussion: https://postgr.es/m/20190423185007.GA27954@alvherre.pgsql
Tom Lane [Wed, 24 Apr 2019 18:15:44 +0000 (14:15 -0400)]
Fix some minor postmaster-state-machine issues.
In sigusr1_handler, don't ignore PMSIGNAL_ADVANCE_STATE_MACHINE based
on pmState. The restriction is unnecessary (PostmasterStateMachine
should work in any state), not future-proof (since it makes too many
assumptions about why the signal might be sent), and broken even today
because a race condition can make it necessary to respond to the signal
in PM_WAIT_READONLY state. The race condition seems unlikely, but
if it did happen, a hot-standby postmaster could fail to shut down
after receiving a smart-shutdown request.
In MaybeStartWalReceiver, don't clear the WalReceiverRequested flag
if the fork attempt fails. Leaving it set allows us to try
again in future iterations of the postmaster idle loop. (The startup
process would eventually send a fresh request signal, but this change
may allow us to retry the fork sooner.)
Remove an obsolete comment and unnecessary test in
PostmasterStateMachine's handling of PM_SHUTDOWN_2 state. It's not
possible to have a live walreceiver in that state, and AFAICT has not
been possible since commit 5e85315ea. This isn't a live bug, but the
false comment is quite confusing to readers.
In passing, rearrange sigusr1_handler's CheckPromoteSignal tests so that
we don't uselessly perform stat() calls that we're going to ignore the
results of.
Add some comments clarifying the behavior of MaybeStartWalReceiver;
I very nearly rearranged it in a way that'd reintroduce the race
condition fixed in e5d494d78. Mea culpa for not commenting that
properly at the time.
Back-patch to all supported branches. The PMSIGNAL_ADVANCE_STATE_MACHINE
change is the only one of even minor significance, but we might as well
keep this code in sync across branches.
postgres_fdw: Fix incorrect handling of row movement for remote partitions.
Commit 3d956d9562 added support for update row movement in postgres_fdw.
This patch fixes the following issues introduced by that commit:
* When a remote partition chosen to insert routed rows into was also an
UPDATE subplan target rel that would be updated later, the UPDATE that
used a direct modification plan modified those routed rows incorrectly
because those routed rows were visible to the later UPDATE command.
The right fix for this would be to have some way in postgres_fdw in
which the later UPDATE command ignores those routed rows, but it seems
hard to do so with the current infrastructure. For now throw an error
in that case.
* When a remote partition chosen to insert routed rows into was also an
UPDATE subplan target rel, fmstate created for the UPDATE that used a
non-direct modification plan was mistakenly overridden by another
fmstate created for inserting those routed rows into the partition.
This caused 1) server crash when the partition would be updated later,
and 2) resource leak when the partition had been already updated. To
avoid that, adjust the treatment of the fmstate for the inserting. As
for #1, since we would also have the incorrectness issue as mentioned
above, error out in that case as well.
Update the docs to mention that postgres_fdw currently does not handle
the case where a remote partition chosen to insert a routed row into is
also an UPDATE subplan target rel that will be updated later.
Author: Amit Langote and Etsuro Fujita Reviewed-by: Amit Langote
Backpatch-through: 11 where row movement in postgres_fdw was added
Discussion: https://postgr.es/m/21e7eaa4-0d4d-20c2-a1f7-c7e96f4ce440@lab.ntt.co.jp
Tom Lane [Tue, 23 Apr 2019 22:51:31 +0000 (18:51 -0400)]
Repair assorted issues in locale data extraction.
cache_locale_time (extraction of LC_TIME-related info) had never been
taught the lessons we previously learned about extraction of info related
to LC_MONETARY and LC_NUMERIC. Specifically, commit 95a777c61 taught
PGLC_localeconv() that data coming out of localeconv() was in an encoding
determined by the relevant locale, but we didn't realize that there's a
similar issue with strftime(). And commit a4930e7ca hardened
PGLC_localeconv() against errors occurring partway through, but failed
to do likewise for cache_locale_time(). So, rearrange the latter
function to perform encoding conversion and not risk failure while
it's got the locales set to temporary values.
This time around I also changed PGLC_localeconv() to treat it as FATAL
if it can't restore the previous settings of the locale values. There
is no reason (except possibly OOM) for that to fail, and proceeding with
the wrong locale values seems like a seriously bad idea --- especially
on Windows where we have to also temporarily change LC_CTYPE. Also,
protect against the possibility that we can't identify the codeset
reported for LC_MONETARY or LC_NUMERIC; rather than just failing,
try to validate the data without conversion.
The user-visible symptom this fixes is that if LC_TIME is set to a locale
name that implies an encoding different from the database encoding,
non-ASCII localized day and month names would be retrieved in the wrong
encoding, leading to either unexpected encoding-conversion error reports
or wrong output from to_char(). The other possible failure modes are
unlikely enough that we've not seen reports of them, AFAIK.
The encoding conversion problems do not manifest on Windows, since
we'd already created special-case code to handle that issue there.
Per report from Juan José Santamaría Flecha. Back-patch to all
supported versions.
Michael Paquier [Tue, 23 Apr 2019 06:43:32 +0000 (15:43 +0900)]
Fix detection of passwords hashed with MD5 or SCRAM-SHA-256
This commit fixes a couple of issues related to the way password
verifiers hashed with MD5 or SCRAM-SHA-256 are detected, leading to
being able to store in catalogs passwords which do not follow the
supported hash formats:
- A MD5-hashed entry was checked based on if its header uses "md5" and
if the string length matches what is expected. Unfortunately the code
never checked if the hash only used hexadecimal characters, as reported
by Tom Lane.
- A SCRAM-hashed entry was checked based on only its header, which
should be "SCRAM-SHA-256$", but it never checked for any fields
afterwards, as reported by Jonathan Katz.
Backpatch down to v10, which is where SCRAM has been introduced, and
where password verifiers in plain format have been removed.
Author: Jonathan Katz Reviewed-by: Tom Lane, Michael Paquier
Discussion: https://postgr.es/m/016deb6b-1f0a-8e9f-1833-a8675b170aa9@postgresql.org
Backpatch-through: 10
Fix documentation of pg_start_backup and pg_stop_backup functions.
This commit adds the description that "non-exclusive" pg_start_backup
and pg_stop_backup can be executed even during recovery. Previously
it was wrongly documented that those functions are not allowed to be
executed during recovery.
Back-patch to 9.6 where non-exclusive backup API was added.
Tom Lane [Fri, 19 Apr 2019 15:20:37 +0000 (11:20 -0400)]
Fix problems with auto-held portals.
HoldPinnedPortals() did things in the wrong order: it must not mark
a portal autoHeld until it's been successfully held. Otherwise,
a failure while persisting the portal results in a server crash
because we think the portal is in a good state when it's not.
Also add a check that portal->status is READY before attempting to
hold a pinned portal. We have such a check before the only other
use of HoldPortal(), so it seems unwise not to check it here.
Lastly, rethink the responsibility for where to call HoldPinnedPortals.
The comment for it imagined that it was optional for any individual PL
to call it or not, but that cannot be the case: if some outer level of
procedure has a pinned portal, failing to persist it when an inner
procedure commits is going to be trouble. Let's have SPI do it instead
of the individual PLs. That's not a complete solution, since in theory
a PL might not be using SPI to perform commit/rollback, but such a PL
is going to have to be aware of lots of related requirements anyway.
(This change doesn't cause an API break for any external PLs that might
be calling HoldPinnedPortals per the old regime, because calling it
twice during a commit or rollback sequence won't hurt.)
Per bug #15703 from Julian Schauder. Back-patch to v11 where this code
came in.
Peter Eisentraut [Tue, 16 Apr 2019 08:37:44 +0000 (10:37 +0200)]
Fix handling of temp and unlogged tables in FOR ALL TABLES publications
If a FOR ALL TABLES publication exists, temporary and unlogged tables
are ignored for publishing changes. But CheckCmdReplicaIdentity()
would still check in that case that such a table has a replica
identity set before accepting updates. To fix, have
GetRelationPublicationActions() return that such a table publishes no
actions.
Bruce Momjian [Wed, 17 Apr 2019 22:12:10 +0000 (18:12 -0400)]
postgresql.conf.sample: add proper defaults for include actions
Previously, include actions include_dir, include_if_exists, and include
listed commented-out values which were not the defaults, which is
inconsistent with other entries. Instead, replace them with '', which
is the default value.
Bruce Momjian [Wed, 17 Apr 2019 22:01:02 +0000 (18:01 -0400)]
docs: clarify pg_upgrade's recovery behavior
The previous paragraph trying to explain --check, --link, and no --link
modes and the various points of failure was too complex. Instead, use
bullet lists and sublists.
Reported-by: Daniel Gustafsson
Discussion: https://postgr.es/m/qtqiv7hI87s_Xvz5ZXHCaH-1-_AZGpIDJowzlRjF3-AbCr3RhSNydM_JCuJ8DE4WZozrtxhIWmyYTbv0syKyfGB6cYMQitp9yN-NZMm-oAo=@yesql.se
projects / postgresql / log