Commit 4aec4989 reorganized the order
of operations here so that we no longer increment the number of "extra
waits" before locking the semaphore, but it did not change the
starting value of extraWaits from 0 to -1 to compensate.  In the worst
case, this could leak a semaphore count, but that seems to be unlikely
in practice.

Discussion: http://postgr.es/m/CAA4eK1JyVqXiMba+-a589Rk0pyHsyKkGxeumVKjU6Y74hdrVLQ@mail.gmail.com

Amit Kapila, per an off-list report by Dilip Kumar.  Reviewed by me.

With the old code, a backend that read pg_stat_activity without ever
having executed a parallel query might see a backend in the midst of
executing one waiting on a DSA LWLock, resulting in a crash.  The
solution is for backends to register the tranche at startup time, not
the first time a parallel query is executed.

Report by Andreas Seltenreich.  Patch by me, reviewed by Thomas Munro.

array_fill(..., array[0]) produced an empty array, which is probably
what users expect, but it was a one-dimensional zero-length array
which is not our standard representation of empty arrays.  Also, for
no very good reason, it rejected empty input arrays; that case should
be allowed and produce an empty output array.

In passing, remove the restriction that the input array(s) have lower
bound 1.  That seems rather pointless, and it would have needed extra
complexity to make the check deal with empty input arrays.

Per bug #14487 from Andrew Gierth.  It's been broken all along, so
back-patch to all supported branches.

Discussion: https://postgr.es/m/20170105152156.10135.64195@wrigleys.postgresql.org

Small number of fixes to perl docs for TAP tests.
Plus two comments that use "xlog" rather than WAL

Michael Paquier

Inheritance operations must treat the OID column, if any, much like
regular user columns.  But MergeAttributesIntoExisting() neglected to
do that, leading to weird results after a table with OIDs is associated
to a parent with OIDs via ALTER TABLE ... INHERIT.

Report and patch by Amit Langote, reviewed by Ashutosh Bapat, some
adjustments by me.  It's been broken all along, so back-patch to
all supported branches.

Discussion: https://postgr.es/m/cb13cfe7-a48c-5720-c383-bb843ab28298@lab.ntt.co.jp

Be more clear that we represent timestamps in microseconds when
integer timestamps are used, and in fractional seconds when
floating-point timestamps are used.

Discussion: http://postgr.es/m/20161212135045.GB15488@e733.localdomain

Report by Alexander Alekseev.  Wording by me with a suggestion
from Tom Lane.

Michael Paquier, per Coverity.

RelationGetPartitionQual() and generate_partition_qual() are always
called with recurse = true, so we don't need an argument for that.

Extracted by me from a larger patch by Amit Langote.

After a tuple is routed to a partition, it has been converted from the
root table's row type to the partition's row type.  ExecConstraints
needs to report the failure using the original tuple and the parent's
tuple descriptor rather than the ones for the selected partition.

Amit Langote

Craig Ringer, reviewed by Euler Taveira and Naoki Okano

Allow pg_recvlogical to specify an ending LSN, complementing
the existing -—startpos=LSN option.

Craig Ringer, reviewed by Euler Taveira and Naoki Okano

configure can only probe the existence of gcc intrinsics, not how well
they're implemented, and unfortunately the answer is sometimes "badly".
In particular we've found that multiple compilers fail to implement
char-width __sync_lock_test_and_set() correctly on PPC; and even a correct
implementation would necessarily be pretty inefficient, since that hardware
has only a word-wide primitive to work with.

Given the knowledge we've accumulated in s_lock.h, it appears that it's
best to rely on int-width TAS operations on most non-Intel architectures.
Hence, pick int not char when both are nominally available to us in
generic-gcc.h (note that that code is not used for x86[_64]).

Back-patch to fix regression test failures on FreeBSD/PPC.  Ordinarily
back-patching a change like this would be verboten because of ABI breakage.
But since pg_atomic_flag is not yet used in any Postgres data structure,
there's no ABI to break.  It seems safer to back-patch to avoid possible
gotchas, if someday we do back-patch something that uses pg_atomic_flag.

Discussion: https://postgr.es/m/25414.1483076673@sss.pgh.pa.us

Commit 2ac3ef7a added a TupleTapleSlot
for partition tuple slot to EState (es_partition_tuple_slot) but it's
more logical to have it as part of ModifyTableState
(mt_partition_tuple_slot) and CopyState (partition_tuple_slot).

Discussion: http://postgr.es/m/1bd459d9-4c0c-197a-346e-e5e59e217d97@lab.ntt.co.jp

Amit Langote, per a gripe from me

Leave OpenSSL's default passphrase collection callback in place during
the first call of secure_initialize() in server startup.  Although that
doesn't work terribly well in daemon contexts, some people feel we should
not break it for anyone who was successfully using it before.  We still
block passphrase demands during SIGHUP, meaning that you can't adjust SSL
configuration on-the-fly if you used a passphrase, but this is no worse
than what it was before commit de41869b.  And we block passphrase demands
during EXEC_BACKEND reloads; that behavior wasn't useful either, but at
least now it's documented.

Tweak some related log messages for more readability, and avoid issuing
essentially duplicate messages about reload failure caused by a passphrase.

Discussion: https://postgr.es/m/29982.1483412575@sss.pgh.pa.us

The typical size of an LWLock is now 16 bytes even on 64-bit platforms,
and the size of slock_t is now irrelevant.  But pg_atomic_uint32 can
(perhaps surprisingly) still be larger than 4 bytes, so there's still
some marginal point to allowing LWLOCK_MINIMAL_SIZE == 64.

Commit 008608b9 made the changes
that led to the need for these updates.

Add new tests for physical repl slots and hot standby feedback.

Craig Ringer, reviewed by Aleksander Alekseev and Simon Riggs

Add methods to the core test framework PostgresNode.pm to allow us to
test that standby nodes have caught up with the master, as well as
basic LSN handling.  Used in tests recovery/t/001_stream_rep.pl and
recovery/t/004_timeline_switch.pl

Craig Ringer, reviewed by Aleksander Alekseev and Simon Riggs

The purpose of the test was to check access to the sequence relation on
a hot standby, so change the test to read a different column from the
sequence, instead of just reading the catalog.

From: Andreas Karlsson <andreas@proxel.se>

These files are deleted but not yet gone from the filesystem. Operations
on them will return ERROR_DELETE_PENDING.

With this we start treating that as ENOENT, meaning files does not
exist (which is the state it will soon reach). This should be safe in
every case except when we try to recreate a file with exactly the same
name. This is an operation that PostgreSQL does very seldom, so
hopefully that won't happen much -- and even if it does, this treatment
should be no worse than treating it as an unhandled error.

We've been un able to reproduce the bug reliably, so pushing this to
master to get buildfarm coverage and other testing. Once it's proven to
be stable, it should be considered for backpatching.

Discussion: https://postgr.es/m/20160712083220.1426.58667%40wrigleys.postgresql.org

Patch by me and Michael Paquier

Since streaming is now supported for all output formats, make this the
default as this is what most people want.

To get the old behavior, the parameter -X none can be specified to turn
it off.

This also removes the parameter -x for fetch, now requiring -X fetch to
be specified to use that.

Reviewed by Vladimir Rusinov, Michael Paquier and Simon Riggs

Backpatch-through: certain files through 9.2

OpenSSL's default behavior when loading a passphrase-protected key file
is to open /dev/tty and demand the password from there.  It was kinda
sorta okay to allow that to happen at server start, but really that was
never workable in standard daemon environments.  And it was a complete
fail on Windows, where the same thing would happen at every backend launch.
Yesterday's commit de41869b put the final nail in the coffin by causing
that to happen at every SIGHUP; even if you've still got a terminal acting
as the server's TTY, having the postmaster freeze until you enter the
passphrase again isn't acceptable.

Hence, override the default behavior with a callback that returns an empty
string, ensuring failure.  Change the documentation to say that you can't
have a passphrase-protected server key, period.

If we can think of a production-grade way of collecting a passphrase from
somewhere, we might do that once at server startup and use this callback
to feed it to OpenSSL, but it's far from clear that anyone cares enough
to invest that much work in the feature.  The lack of complaints about
the existing fractionally-baked behavior suggests nobody's using it anyway.

Discussion: https://postgr.es/m/29982.1483412575@sss.pgh.pa.us

From: Kuntal Ghosh <kuntalghosh.2007@gmail.com>

That was written when we still had "crypt" authentication, and it was
referring to the fact that an older client might support "crypt"
authentication but not "md5". But we haven't supported "crypt" for years.
(As soon as we add a new authentication mechanism that doesn't work with
MD5 hashes, we'll need a similar notice again. But this text as it's worded
now is just wrong.)

Backpatch to all supported versions.

Discussion: https://www.postgresql.org/message-id/9a7263eb-0980-2072-4424-440bb2513dc7@iki.fi

It is no longer necessary to restart the server to enable, disable,
or reconfigure SSL.  Instead, we just create a new SSL_CTX struct
(by re-reading all relevant files) whenever we get SIGHUP.  Testing
shows that this is fast enough that it shouldn't be a problem.

In conjunction with that, downgrade the logic that complains about
pg_hba.conf "hostssl" lines when SSL isn't active: now that's just
a warning condition not an error.

An issue that still needs to be addressed is what shall we do with
passphrase-protected server keys?  As this stands, the server would
demand the passphrase again on every SIGHUP, which is certainly
impractical.  But the case was only barely supported before, so that
does not seem a sufficient reason to hold up committing this patch.

Andreas Karlsson, reviewed by Michael Banck and Michael Paquier

Discussion: https://postgr.es/m/556A6E8A.9030400@proxel.se

The advantage of clock_gettime() is that the API allows the result to
be precise to nanoseconds, not just microseconds as in gettimeofday().
Now that it's routinely possible to do tens of plan node executions
in 1us, we really need more precision than gettimeofday() can offer
for EXPLAIN ANALYZE to accumulate statistics with.

Some research shows that clock_gettime() is available on pretty nearly
every modern Unix-ish platform, and as far as I have been able to test,
it has about the same execution time as gettimeofday(), so there's no
loss in switching over.  (By the same token, this doesn't do anything
to fix the fact that we really wish clock readings were faster.  But
there's enough win here to justify changing anyway.)

A small side benefit is that on most platforms, we can use CLOCK_MONOTONIC
instead of CLOCK_REALTIME and thereby render EXPLAIN impervious to
concurrent resets of the system clock.  (This means that code must not
assume that the contents of struct instr_time have any well-defined
interpretation as timestamps, but really that was true before.)

Some platforms offer nonstandard clock IDs that might be of interest.
This patch knows we should use CLOCK_MONOTONIC_RAW on macOS, because it
provides more precision and is faster to read than their CLOCK_MONOTONIC.
If there turn out to be many more cases where we need special rules, it
might be appropriate to handle the selection of clock ID in configure,
but for the moment that doesn't seem worth the trouble.

Discussion: https://postgr.es/m/31856.1400021891@sss.pgh.pa.us

For aggregated logging, pg_bench supposed that printing the integer part of
INSTR_TIME_GET_DOUBLE() would produce a Unix timestamp.  That was already
broken on Windows, and it's about to get broken on most other platforms as
well.  As in commit 74baa1e3, we can remove the entanglement at the price
of one extra syscall per transaction; though here it seems more convenient
to use time(NULL) instead of gettimeofday(), since we only need
integral-second precision.

I took the time to do some wordsmithing on the documentation about
pgbench's logging features, too.

Discussion: https://postgr.es/m/8837.1483216839@sss.pgh.pa.us

This code was presuming undue familiarity with the contents of the
instr_time struct.  That was already broken on Windows, and it's about
to get broken on most other platforms as well.  The simplest solution
that preserves the current output definition is to just do our own
gettimeofday() call here.  Realistically, the extra cost is probably
negligible in comparison to everything else that's going on in a
pgbench transaction, so it's not worth sweating over.

On Windows, the precision delivered by gettimeofday() is lower than
one could wish, but this is still a big improvement over printing
zeroes, as the code did before.

Discussion: https://postgr.es/m/8837.1483216839@sss.pgh.pa.us

Commit 2ac3ef7a added a query with an underdetermined output row order;
it has failed multiple times in the buildfarm since then.  Add an ORDER BY
to fix.  Also, don't rely on a DROP CASCADE to drop in a well-determined
order; that hasn't failed yet but I don't trust it much, and we're not
saving any typing by using CASCADE anyway.

Commit f0e44751 added new node tags at a place in the tag numbering
where there was no daylight left before the next hard-coded number,
resulting in some duplicate tag assignments.  This doesn't seem to have
caused any big problem so far, but it's surely trouble waiting to happen.

We could adjust the manually assigned breakpoints to make more room,
but that just leaves the same hazard waiting to strike again in future.
What seems like a better idea is to get rid of the manual assignments
and leave NodeTags to be automatically assigned, consecutively from one
on up.  This means that any change in the tag list forces a backend-wide
recompile, but realistically that's usually needed anyway.

Discussion: https://postgr.es/m/29670.1482942811@sss.pgh.pa.us

There is no need to use abbreviations here, so just write it out for
consistency.

Most code was casting this through a generic Node.  By declaring
everything as RoleSpec appropriately, we can remove a bunch of casts and
ad-hoc node type checking.
Reviewed-by: Alvaro Herrera <alvherre@2ndquadrant.com>

interval_transform() contained two separate bugs that caused it to
sometimes mistakenly decide that a cast from interval to restricted
interval is a no-op and throw it away.

First, it was wrong to rely on dt.h's field type macros to have an
ordering consistent with the field's significance; in one case they do
not.  This led to mistakenly treating YEAR as less significant than MONTH,
so that a cast from INTERVAL MONTH to INTERVAL YEAR was incorrectly
discarded.

Second, fls(1<<k) produces k+1 not k, so comparing its output directly
to SECOND was wrong.  This led to supposing that a cast to INTERVAL
MINUTE was really a cast to INTERVAL SECOND and so could be discarded.

To fix, get rid of the use of fls(), and make a function based on
intervaltypmodout to produce a field ID code adapted to the need here.

Per bug #14479 from Piotr Stefaniak.  Back-patch to 9.2 where transform
functions were introduced, because this code was born broken.

Discussion: https://postgr.es/m/20161227172307.10135.7747@wrigleys.postgresql.org

In addition to space accounted for by tuple_len, dead_tuple_len and
free_space, the table_len includes page overhead, the item pointers
table and padding bytes.

Backpatch to live branches.

In 56c7d8d4 the behavior to keep .partial segments around
(considered corrupt) in case an connection failure occurs was
accidentally removed. This would lead to an incomplete segment
being considered complete.

Author: Michael Paquier

Erik Rijkers