The multirange_get_range() function fails when two boundaries of the same
range have different alignments. Fix that by adding proper pointer alignment.
Reported-by: Alexander Lakhin
Discussion: https://postgr.es/m/17300-dced2d01ddeb1f2f%40postgresql.org
Backpatch-through: 14
This commit improves the description of some WAL records for the
Transaction RMGR:
- Track remote_apply for a transaction commit. This GUC is
user-settable, so this information can be useful for debugging.
- Add replication origin information for PREPARE TRANSACTION, with the
origin ID, LSN and timestamp
- Same as above, for ROLLBACK PREPARED.
This impacts the format of pg_waldump or anything using these
description routines, so no backpatch is done.
Author: Masahiko Sawada, Michael Paquier
Discussion: https://postgr.es/m/CAD21AoD2dJfgsdxk4_KciAZMZQoUiCvmV9sDpp8ZuKLtKCNXaA@mail.gmail.com
Historically we've put type "char" into the S (String) typcategory,
although calling it a string is a stretch considering it can only
store one byte. (In our actual usage, it's more like an enum.)
This choice now seems wrong in view of the special heuristics
that parse_func.c and parse_coerce.c have for TYPCATEGORY_STRING:
it's not a great idea for "char" to have those preferential casting
behaviors.
Worse than that, recent patches inventing special-purpose types
like pg_node_tree have assigned typcategory S to those types,
meaning they also get preferential casting treatment that's designed
on the assumption that they can hold arbitrary text.
To fix, invent a new category TYPCATEGORY_INTERNAL for internal-use
types, and assign that to all these types. I used code 'Z' for
lack of a better idea ('I' was already taken).
This change breaks one query in psql/describe.c, which now needs to
explicitly cast a catalog "char" column to text before concatenating
it with an undecorated literal. Also, a test case in contrib/citext
now needs an explicit cast to convert citext to "char". Since the
point of this change is to not have "char" be a surprisingly-available
cast target, these breakages seem OK.
Per report from Ian Campbell.
Discussion: https://postgr.es/m/2216388.1638480141@sss.pgh.pa.us
The test added by 5753d4ee32 relies on statistics collector, and so it
may occasionally fail when the UDP packet gets lost. Some machines may
be susceptible to this, probably depending on load etc.
Move the test to stats.sql, which is known to already have this issue
and people know to ignore it.
Reported-by: Justin Pryzby
Discussion: https://postgr.es/m/CAFp7QwpMRGcDAQumN7onN9HjrJ3u4X3ZRXdGFT0K5G2JWvnbWg%40mail.gmail.com
1. Update our open() wrapper to check for NT's STATUS_DELETE_PENDING
and translate it to Unix-like errors. This is done with
RtlGetLastNtStatus(), which is dynamically loaded from ntdll. A new
file win32ntdll.c centralizes lookup of NT functions, in case we decide
to add more in the future.
2. Remove non-working code that was trying to do something similar for
stat(), and just reuse the open() wrapper code. As a side effect,
stat() also gains resilience against "sharing violation" errors.
3. Since stat() is used very early in process startup, remove the
requirement that the Win32 signal event has been created before
pgwin32_open_handle() is reached. Instead, teach pg_usleep() to fall
back to a non-interruptible sleep if reached before the signal event is
available.
This could be back-patched, but for now it's in master only. The
problem has apparently been with us for a long time and generated only a
few complaints. Proposed patches trigger it more often, which led to
this investigation and fix.
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Alexander Lakhin <exclusion@gmail.com>
Reviewed-by: Juan José Santamaría Flecha <juanjo.santamaria@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKGJz_pZTF9mckn6XgSv69%2BjGwdgLkxZ6b3NWGLBCVjqUZA%40mail.gmail.com
We publish the child table's data twice for a publication that has both
child and parent tables and is published with publish_via_partition_root
as true. This happens because subscribers will initiate synchronization
using both parent and child tables, since it gets both as separate tables
in the initial table list.
Ensure that pg_publication_tables returns only parent tables in such
cases.
Author: Hou Zhijie
Reviewed-by: Greg Nancarrow, Amit Langote, Vignesh C, Amit Kapila
Backpatch-through: 13
Discussion: https://postgr.es/m/OS0PR01MB57167F45D481F78CDC5986F794B99@OS0PR01MB5716.jpnprd01.prod.outlook.com
The term "super-exclusive lock" is a synonym for "buffer cleanup lock"
that first appeared in nbtree many years ago. Standardize things by
consistently using the term cleanup lock. This finishes work started by
commit 276db875.
There is no good reason to have two terms. But there is a good reason
to only have one: to avoid confusion around why VACUUM acquires a full
cleanup lock (not just an ordinary exclusive lock) in index AMs, during
ambulkdelete calls. This has nothing to do with protecting the physical
index data structure itself. It is needed to implement a locking
protocol that ensures that TIDs pointing to the heap/table structure
cannot get marked for recycling by VACUUM before it is safe (which is
somewhat similar to how VACUUM uses cleanup locks during its first heap
pass). Note that it isn't strictly necessary for index AMs to implement
this locking protocol -- several index AMs use an MVCC snapshot as their
sole interlock to prevent unsafe TID recycling.
In passing, update the nbtree README. Cleanly separate discussion of
the aforementioned index vacuuming locking protocol from discussion of
the "drop leaf page pin" optimization added by commit 2ed5b87f. We now
structure discussion of the latter by describing how individual index
scans may safely opt out of applying the standard locking protocol (and
so can avoid blocking progress by VACUUM). Also document why the
optimization is not safe to apply during nbtree index-only scans.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-WzngHgQa92tz6NQihf4nxJwRzCV36yMJO_i8dS+2mgEVKw@mail.gmail.com
Discussion: https://postgr.es/m/CAH2-WzkHPgsBBvGWjz=8PjNhDefy7XRkDKiT5NxMs-n5ZCf2dA@mail.gmail.com
Extend the foreign key ON DELETE actions SET NULL and SET DEFAULT by
allowing the specification of a column list, like
CREATE TABLE posts (
...
FOREIGN KEY (tenant_id, author_id) REFERENCES users ON DELETE SET NULL (author_id)
);
If a column list is specified, only those columns are set to
null/default, instead of all the columns in the foreign-key
constraint.
This is useful for multitenant or sharded schemas, where the tenant or
shard ID is included in the primary key of all tables but shouldn't be
set to null.
Author: Paul Martinez <paulmtz@google.com>
Discussion: https://www.postgresql.org/message-id/flat/CACqFVBZQyMYJV=njbSMxf+rbDHpx=W=B7AEaMKn8dWn9OZJY7w@mail.gmail.com
We show duplicate values for child tables in publications that have both
child and parent tables and are published with publish_via_partition_root
as false which is not what the user would expect.
We decided not to backpatch this as there is no user complaint about this
and it doesn't seem to be a critical issue.
Author: Hou Zhijie
Reviewed-by: Bharath Rupireddy, Amit Langote, Amit Kapila
Discussion: https://postgr.es/m/OS0PR01MB5716E97F00732B52DC2BBC2594989@OS0PR01MB5716.jpnprd01.prod.outlook.com
This simplifies the code so as it is not necessary anymore for the
caller of parse_subscription_options() to zero SubOpts, holding a
bitmaps of the provided options as well as the default/parsed option
values. This also simplifies some checks related to the options
supported by a command when checking for incompatibilities.
While on it, the errors generated for unsupported combinations with
"slot_name = NONE" are reordered. This may generate a different errors
compared to the previous major versions, but users have to go through
all those errors to get a correct command in this case when using
incorrect values for options "enabled" and "create\slot", so at the end
the resulting command would remain the same.
Author: Peter Smith
Reviewed-by: Nathan Bossart
Discussion: https://postgr.es/m/CAHut+PtXHfLgLHDDJ8ZN5f5Be_37mJoxpEsRg8LNmm4XCr06Rw@mail.gmail.com
REINDEX CONCURRENTLY run on a toast index or a toast relation could
corrupt the target indexes rebuilt, as a backend running in parallel
that manipulates toast values would directly release the lock on the
toast relation when its local operation is done, rather than releasing
the lock once the transaction that manipulated the toast values
committed.
The fix done here is simple: we now hold a ROW EXCLUSIVE lock on the
toast relation when saving or deleting a toast value until the
transaction working on them is committed, so as a concurrent reindex
happening in parallel would be able to wait for any activity and see any
new rows inserted (or deleted).
An isolation test is added to check after the case fixed here, which is
a bit fancy by design as it relies on allow_system_table_mods to rename
the toast table and its index to fixed names. This way, it is possible
to reindex them directly without any dependency on the OID of the
underlying relation. Note that this could not use a DO block either, as
REINDEX CONCURRENTLY cannot be run in a transaction block. The test is
backpatched down to 13, where it is possible, thanks to c4a7a39, to use
allow_system_table_mods in a test suite.
Reported-by: Alexey Ermakov
Analyzed-by: Andres Freund, Noah Misch
Author: Michael Paquier
Reviewed-by: Nathan Bossart
Discussion: https://postgr.es/m/17268-d2fb426e0895abd4@postgresql.org
Backpatch-through: 12
Certain settings from configuration or the Makefile infrastructure are
used by the TAP tests, but were not being set up by vcregress.pl. This
remedies those omissions. This should increase test coverage, especially
on the buildfarm.
Reviewed by Noah Misch
Discussion: https://postgr.es/m/17093da5-e40d-8335-d53a-2bd803fc38b0@dunslane.net
Backpatch to all live branches.
Further experimentation shows that commit 6051857fc is not sufficient
when using (some versions of?) OpenSSL. The reason is obscure, but
calling shutdown(socket, SD_SEND) improves matters.
Per testing by Andrew Dunstan and Alexander Lakhin.
Back-patch as before.
Discussion: https://postgr.es/m/af5e0bf3-6a61-bb97-6cba-061ddf22ff6b@dunslane.net
In parallel mode, pg_dump tries to order the table-data-dumping
jobs with the largest tables first. However, it was only
consulting the pg_class.relpages value to determine table size.
This ignores TOAST data, and so we could make poor scheduling
decisions in cases where some large tables are mostly TOASTed
data while others have very little. To fix, add in the relpages
value for the TOAST table as well.
This patch also fixes a potential integer-overflow issue that
could result in poor scheduling on machines where off_t is
only 32 bits wide. Such platforms are probably extinct in the
wild, but we do still nominally support them, so repair.
Per complaint from Hans Buschmann.
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
For objects such as functions, pg_dump issues the same secondary
data-collection query against each object to be dumped. This can't
readily be refactored to avoid the repetitive queries, but we can
PREPARE these queries to reduce planning costs.
This patch applies the idea to functions, aggregates, operators, and
data types. While it could be carried further, the remaining sorts of
objects aren't likely to appear in typical databases enough times to
be worth worrying over. Moreover, doing the PREPARE is likely to be a
net loss if there aren't at least some dozens of objects to apply the
prepared query to.
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
Instead of issuing a secondary data-collection query against each
table to be dumped, issue just one query, with a WHERE clause
restricting it to be applied to only the tables we intend to dump.
Likewise for indexes, constraints, and triggers. This greatly
reduces the number of queries needed to dump a database containing
many tables. It might seem that WHERE clauses listing many target
OIDs could be inefficient, but at least on recent server versions
this provides a very substantial speedup.
(In principle the same thing could be done with other object types
such as functions; but that would require significant refactoring
of pg_dump, so those will be tackled in a different way in a
following patch.)
The new WHERE clauses depend on the unnest() function, which is
only present in 8.4 and above. We could implement them differently
for older servers, but there is an ongoing discussion that will
probably result in dropping pg_dump support for servers before 9.2,
so that seems like it'd be wasted work. For now, just bump the
server version check to require >= 8.4, without stopping to remove
any of the code that's thereby rendered dead. We'll mop that
situation up soon.
Patch by me, based on an idea from Andres Freund.
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
Avoid calling pg_get_partkeydef(), pg_get_expr(relpartbound),
and regtypeout until we have lock on the relevant tables.
The existing coding is at serious risk of failure if there
are any concurrent DROP TABLE commands going on --- including
drops of other sessions' temp tables.
Arguably this is a bug fix that should be back-patched, but it's
moderately invasive and we've not had all that many complaints
about such failures. Let's just put it in HEAD for now.
Discussion: https://postgr.es/m/2273648.1634764485@sss.pgh.pa.us
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
Throw away most of the existing logic for this, as it was very
inefficient thanks to expensive sub-selects executed to collect
ACL data that we very possibly would have no interest in dumping.
Reduce the ACL handling in the initial per-object-type queries
to be just collection of the catalog ACL fields, as it was
originally. Fetch pg_init_privs data separately in a single
scan of that catalog, and do the merging calculations on the
client side. Remove the separate code path used for pre-9.6
source servers; there is no good reason to treat them differently
from newer servers that happen to have empty pg_init_privs.
Discussion: https://postgr.es/m/2273648.1634764485@sss.pgh.pa.us
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
Split the DumpableObject.dump bitmask field into separate bitmasks
tracking which components are requested to be dumped (in the
existing "dump" field) and which components exist for the particular
object (in the new "components" field). This gets rid of some
klugy and easily-broken logic that involved setting bits and later
clearing them. More importantly, it restores the originally intended
behavior that pg_dump's secondary data-gathering queries should not
be executed for objects we have no interest in dumping. That
optimization got broken when the dump flag was turned into a bitmask,
because irrelevant bits tended to remain set in many cases. Since
the "components" field starts from a minimal set of bits and is
added onto as needed, ANDing it with "dump" provides a reliable
indicator of what we actually have to dump, without having to
complicate the logic that manages the request bits. This makes
a significant difference in the number of queries needed when,
for example, there are many functions in extensions.
Discussion: https://postgr.es/m/2273648.1634764485@sss.pgh.pa.us
Discussion: https://postgr.es/m/7d7eb6128f40401d81b3b7a898b6b4de@W2012-02.nidsa.loc
The chr() function used PG_GETARG_UINT32() even though the argument is
declared as (signed) integer. As a result, you can pass negative
arguments to this function and it internally interprets them as
positive. Ultimately ends up being harmless, but it seems wrong, so
fix this and rearrange the internal error checking a bit to
accommodate this.
Another case was in the documentation, where example code used
PG_GETARG_UINT32() with an argument declared as signed integer.
Reviewed-by: Nathan Bossart <bossartn@amazon.com>
Discussion: https://www.postgresql.org/message-id/flat/7e43869b-d412-8f81-30a3-809783edc9a3%40enterprisedb.com
Perl's hex() function complains if its argument contains trailing white
space (or in fact anything other than hex digits), so remove the
offending text.
The temporary path generated in commit c113d8ad5 cannot be passed as-is in
the connection string on Windows since the path delimiting backslashes will
be treated as escape characters. Fix by converting backslash to slash as in
similar path usecases in other tests.
Reported-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20211202195130.e7pprpsx4ell22sp@alap3.anarazel.de
This commit fixes a couple of inconsistencies in the descriptions of
some GUCs, while making their wording more general regarding the units
they rely on.
For most of them, this removes the use of terms like "N seconds" or "N
bytes", which may not apply easily to all the languages these strings
are translated to (from my own experience, this works in French and
English, less in Japanese).
Per debate between the authors listed below.
Author: Justin Pryzby, Michael Paquier
Discussion: https://postgr.es/m/20211129030833.GJ17618@telsasoft.com
It turns out that this is necessary to keep Winsock from dropping any
not-yet-sent data, such as an error message explaining the reason for
process termination. It's pretty weird that the implicit close done
by the kernel acts differently from an explicit close, but it's hard
to argue with experimental results.
Independently submitted by Alexander Lakhin and Lars Kanis (comments
by me, though). Back-patch to all supported branches.
Discussion: https://postgr.es/m/90b34057-4176-7bb0-0dbb-9822a5f6425b@greiz-reinsdorf.de
Discussion: https://postgr.es/m/16678-253e48d34dc0c376@postgresql.org
Checking for RELKIND_MATVIEW was forgotten in
guessConstraintInheritance(). This isn't a live problem, since it is
checked in flagInhTables() which relkinds can have parents, and those
entries will have numParents==0 after that. But after discussion it
was felt that this place should be kept consistent with
flagInhTables() and flagInhAttrs().
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://www.postgresql.org/message-id/flat/a574c8f1-9c84-93ad-a9e5-65233d6fc00f@enterprisedb.com
Previously, pg_waldump would not display its statistics summary if it
got interrupted by SIGINT (or say a simple Ctrl+C). It gains with this
commit a signal handler for SIGINT, trapping the signal to exit at the
earliest convenience to allow a display of the stats summary before
exiting. This makes the reports more interactive, similarly to strace
-c.
This new behavior makes the combination of the options --stats and
--follow much more useful, so as the user will get a report for any
invocation of pg_waldump in such a case. Information about the LSN
range of the stats computed is added as a header to the report
displayed.
This implementation comes from a suggestion by Álvaro Herrera and
myself, following a complaint by the author of this patch about --stats
and --follow not being useful together originally.
As documented, this is not supported on Windows, though its support
would be possible by catching the terminal events associated to Ctrl+C,
for example (this may require a more centralized implementation, as
other tools could benefit from a common API).
Author: Bharath Rupireddy
Discussion: https://postgr.es/m/CALj2ACUUx3PcK2z9h0_m7vehreZAUbcmOky9WSEpe8TofhV=PQ@mail.gmail.com
The existing pg_upgrade/test.sh and the buildfarm code have been holding
the same set of SQL queries when doing cross-version upgrade tests to
adapt the objects created by the regression tests before the upgrade
(mostly, incompatible or non-existing objects need to be dropped from
the origin, perhaps re-created).
This moves all those SQL queries into a new, separate, file with a set
of \if clauses to handle the version checks depending on the old version
of the cluster to-be-upgraded.
The long-term plan is to make the buildfarm code re-use this new SQL
file, so as committers are able to fix any compatibility issues in the
tests of pg_upgrade with a refresh of the core code, without having to
poke at the buildfarm client. Note that this is only able to handle the
main regression test suite, and that nothing is done yet for contrib
modules yet (these have more issues like their database names).
A backpatch down to 10 is done, adapting the version checks as this
script needs to be only backward-compatible, so as it becomes possible
to clean up a maximum amount of code within the buildfarm client.
Author: Justin Pryzby, Michael Paquier
Discussion: https://postgr.es/m/20201206180248.GI24052@telsasoft.com
Backpatch-through: 10
The various ALTER OWNER routines tend to leak memory in
CurrentMemoryContext. That's not a problem when they're only called
once per command; but in this usage where we might be touching many
objects, it can amount to a serious memory leak. Fix that by running
each call in a short-lived context.
(DROP OWNED BY likely has a similar issue, except that you'll probably
run out of lock table space before noticing. REASSIGN is worth fixing
since for most non-table object types, it won't take any lock.)
Back-patch to all supported branches. Unfortunately, in the back
branches this helps to only a limited extent, since the sinval message
queue bloats quite a lot in this usage before commit 3aafc030a,
consuming memory more or less comparable to what's actually leaked.
Still, it's clearly a leak with a simple fix, so we might as well fix it.
Justin Pryzby, per report from Guillaume Lelarge
Discussion: https://postgr.es/m/CAECtzeW2DAoioEGBRjR=CzHP6TdL=yosGku8qZxfX9hhtrBB0Q@mail.gmail.com
Readline's meta-# command is supposed to insert a comment marker
at the start of the current line. However, the default marker is
"#" which is entirely unhelpful for SQL. Set it to "-- " instead.
(This setting can still be overridden in one's ~/.inputrc file,
so this change won't affect people who have already taken steps
to make the command useful.)
Discussion: https://postgr.es/m/CAJcOf-cAdMVr7azeYR7nWKsNp7qhORzc84rV6d7m7knG5Hrtsw@mail.gmail.com
If we've not yet collected any non-whitespace, non-comment token for a
new query, flush the current input line to history before reading
another line. This aligns psql's history behavior with the observation
that lines containing only comments are generally not thought of as
being part of the next query. psql's prompting behavior is consistent
with that view, too, since it won't change the prompt until you
enter something that's neither whitespace nor a "--" comment.
Greg Nancarrow, simplified a bit by me
Discussion: https://postgr.es/m/CAJcOf-cAdMVr7azeYR7nWKsNp7qhORzc84rV6d7m7knG5Hrtsw@mail.gmail.com
psql's lexer has historically deleted dash-dash (single-line) comments
from what's collected and sent to the server. This is inconsistent
with what it does for slash-star comments, and people have complained
before that they wish such comments would be captured in the server log.
Undoing the decision completely seems like too big a behavioral change,
however. In particular, comments on lines preceding the start of a
query are generally not thought of as being part of that query.
What we can do to improve the situation is to capture comments that
are clearly *within* a query, that is after the first non-whitespace,
non-comment token but before the query's ending semicolon or backslash
command. This is a nearly trivial code change, and it affects only a
few regression test results.
(It is tempting to try to apply the same rule to slash-star comments.
But it's hard to see how to do that without getting strange history
behavior for comments that cross lines, especially if the user then
starts a new query on the same line as the star-slash. In view of
the lack of complaints, let's leave that case alone.)
Discussion: https://postgr.es/m/CAJcOf-cAdMVr7azeYR7nWKsNp7qhORzc84rV6d7m7knG5Hrtsw@mail.gmail.com
An extension can already de facto reserve a GUC prefix using
EmitWarningsOnPlaceholders(). But this was only checked against
settings that exist at the time the extension is loaded (or the
extension chooses to call this). No diagnostic is given when a SET
command later uses a nonexisting setting with a custom prefix.
With this change, EmitWarningsOnPlaceholders() saves the prefixes it
reserves in a list, and SET checks when it finds a "placeholder"
setting whether it belongs to a reserved prefix and issues a warning
in that case.
Add a regression test that checks the patch using the "plpgsql"
registered prefix.
Author: Florin Irion <florin.irion@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/CA+HEvJDhWuuTpGTJT9Tgbdzm4QS4EzPAwDBScWK18H2Q=FVJFw@mail.gmail.com
Commit c113d8ad50 moved the copying of certificates into a temporary path
for the duration of the tests, instead of using the source tree. This broke
the tests on msys as the absolute path wasn't adapted for the msys platform.
Ensure to convert the path with perl2host before copying and passing in the
connection string.
While there also make certificate copying error handling uniform across all
the test suites.
Discussion: https://postgr.es/m/YacT3tm97xziSUFw@paquier.xyz
The tests didn't considered that an error unrelated to apply changes, e.g.
"replication origin with OID %d is already active ...", could occur on the
table sync worker before starting to copy changes.
To make the test robust we instead need to check the expected error and
the source of error which will be either tablesync or apply worker.
In passing remove the harmless option "streaming = off" from Create
Subscription command as that is anyway the default.
Per buildfarm member sidewinder.
Author: Masahiko Sawada
Reviewed-by: Hou Zhijie, Vignesh C, Amit Kapila
Discussion: https://postgr.es/m/CAD21AoDeScrsHhLyEPYqN3sydg6PxAPVBboK=30xJfUVihNZDA@mail.gmail.com
Discussion: https://postgr.es/m/E1mrtvV-0002Gz-73@gemulon.postgresql.org
If the stat operation on the private key failed, the code assumed it
was due to an ENOENT, which may or may not be true. Extend the check
by printing a different error message on non-ENOENT errors for easier
debugging.
Per suggestion by Tom Lane due to an issue with the fairywren animal
in the buildfarm.
Discussion: https://postgr.es/m/1632478.1638305700@sss.pgh.pa.us
When determining whether an index update may be skipped by using HOT, we
can ignore attributes indexed only by BRIN indexes. There are no index
pointers to individual tuples in BRIN, and the page range summary will
be updated anyway as it relies on visibility info.
This also removes rd_indexattr list, and replaces it with rd_attrsvalid
flag. The list was not used anywhere, and a simple flag is sufficient.
Patch by Josef Simanek, various fixes and improvements by me.
Author: Josef Simanek
Reviewed-by: Tomas Vondra, Alvaro Herrera
Discussion: https://postgr.es/m/CAFp7QwpMRGcDAQumN7onN9HjrJ3u4X3ZRXdGFT0K5G2JWvnbWg%40mail.gmail.com
Alexander Korotkov
Michael Paquier
Tom Lane
Tomas Vondra
Thomas Munro
Amit Kapila
Peter Geoghegan
Andrew Dunstan
Peter Eisentraut
Daniel Gustafsson