Previously server reserved WAL for last two checkpoints,
which used too much disk space for small servers.
Bumps PG_CONTROL_VERSION
Author: Simon Riggs <simon@2ndQuadrant.com>
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Add docs to explain this for other backup mechanisms
Author: David Steele <david@pgmasters.net>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndQuadrant.com> et al
The update path of an INSERT ... ON CONFLICT DO UPDATE requires SELECT
permission on the columns of the arbiter index, but it failed to check
for that in the case of an arbiter specified by constraint name.
In addition, for a table with row level security enabled, it failed to
check updated rows against the table's SELECT policies when the update
path was taken (regardless of how the arbiter index was specified).
Backpatch to 9.5 where ON CONFLICT DO UPDATE and RLS were introduced.
Security: CVE-2017-15099
NSUnLinkModule() doesn't take a bool as second argument but one of set
of specific constants. The numeric values are the same in this case,
but clean it up while we're cleaning up bool use elsewhere.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
There doesn't seem to be any good reason to do the filling of the
itemidbase[] array separately from the first traversal of the pointers.
It's certainly not a win if there are any line pointers with storage,
and even if there aren't, this change doesn't insert code into the part
of the first loop that will be traversed in that case. So let's just
merge the two loops.
Yura Sokolov, reviewed by Claudio Freire
Discussion: https://postgr.es/m/e49befcc6f1d7099834c6fdf5c675a60@postgrespro.ru
btree, hash, and bloom indexes all set up their metapages in standard
format (that is, with pd_lower and pd_upper correctly delimiting the
unused area); but they mostly didn't inform the xlog routines of this.
When calling log_newpage[_buffer], this is bad because it loses the
opportunity to compress unused data out of the WAL record. When
calling XLogRegisterBuffer, it's not such a performance problem because
all of these call sites also use REGBUF_WILL_INIT, preventing an FPI
image from being written. But it's still a good idea to provide the
flag when relevant, because that aids WAL consistency checking.
This completes the project of getting all the in-core index AMs to
handle their metapage WAL operations similarly.
Amit Kapila, reviewed by Michael Paquier
Discussion: https://postgr.es/m/0d273805-0e9e-ec1a-cb84-d4da400b8f85@lab.ntt.co.jp
The previous commit contained a thinko that made a single-range
summarization request process from there to end of table. Fix by
setting the correct end range point. Per buildfarm.
When a publisher table has fewer columns than a subscriber, the update
of a row on the publisher should result in updating of only the columns
in common. The previous coding mistakenly reset the values of
additional columns on the subscriber to NULL because it failed to skip
updates of columns not found in the attribute map.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
If a process is extending a table concurrently with some BRIN
summarization process, it is possible for the latter to miss pages added
by the former because the number of pages is computed ahead of time.
Fix by determining a fresh relation size after inserting the placeholder
tuple: any process that further extends the table concurrently will
update the placeholder tuple, while previous pages will be processed by
the heap scan.
Reported-by: Tomas Vondra
Reviewed-by: Tom Lane
Author: Álvaro Herrera
Discussion: https://postgr.es/m/083d996a-4a8a-0e13-800a-851dd09ad8cc@2ndquadrant.com
Backpatch-to: 9.5
Previously, these index types left the pd_lower field set to the default
SizeOfPageHeaderData, which is really a lie because it ought to point past
whatever space is being used for metadata. The coding accidentally failed
to fail because we never told xlog.c that the metapage is of standard
format --- but that's not very good, because it impedes WAL consistency
checking, and in some cases prevents compression of full-page images.
To fix, ensure that we set pd_lower correctly, not only when creating a
metapage but whenever we write it out (these apparently redundant steps are
needed to cope with pg_upgrade'd indexes that don't yet contain the right
value). This allows telling xlog.c that the page is of standard format.
The WAL consistency check mask functions are made to mask only if pd_lower
appears valid, which I think is likely unnecessary complication, since
any metapage appearing in a v11 WAL stream should contain valid pd_lower.
But it doesn't cost much to be paranoid.
Amit Langote, reviewed by Michael Paquier and Amit Kapila
Discussion: https://postgr.es/m/0d273805-0e9e-ec1a-cb84-d4da400b8f85@lab.ntt.co.jp
In some cases the BRIN code releases lock on an index page, and later
re-acquires lock and tries to check that the tuple it was working on is
still there. That check was a couple bricks shy of a load. It didn't
consider that the page might have turned into a "revmap" page. (The
samepage code path doesn't call brin_getinsertbuffer(), so it isn't
protected by the checks for revmap status there.) It also didn't check
whether the tuple offset was now off the end of the linepointer array.
Since commit 24992c6db the latter case is pretty common, but at least
in principle it could have occurred before that. The net result is
that concurrent updates of a BRIN index could fail with errors like
"invalid index offnum" or "inconsistent range map".
Per report from Tomas Vondra. Back-patch to 9.5, since this code is
substantially the same in all versions containing BRIN.
Discussion: https://postgr.es/m/10d2b9f9-f427-03b8-8ad9-6af4ecacbee9@2ndquadrant.com
For some reason, we have never accounted for either the evaluation cost
or the selectivity of filter conditions attached to Agg and Group nodes
(which, in practice, are always conditions from a HAVING clause).
Applying our regular selectivity logic to post-grouping conditions is a
bit bogus, but it's surely better than taking the selectivity as 1.0.
Perhaps someday the extended-statistics mechanism can be taught to provide
statistics that would help us in getting non-default estimates here.
Per a gripe from Benjamin Coutu. This is surely a bug fix, but I'm
hesitant to back-patch because of the prospect of destabilizing existing
plan choices. Given that it took us this long to notice the bug, it's
probably not hurting too many people in the field.
Discussion: https://postgr.es/m/20968.1509486337@sss.pgh.pa.us
It turns out we misdiagnosed what the real problem was. Revert the
previous changes, because they may have worse consequences going
forward. A better fix is forthcoming.
The simplistic test case is kept, though disabled.
Discussion: https://postgr.es/m/20171102112019.33wb7g5wp4zpjelu@alap3.anarazel.de
If we don't have to return any columns from heap tuples, and there's
no need to recheck qual conditions, and the heap page is all-visible,
then we can skip fetching the heap page altogether.
Skip prefetching pages too, when possible, on the assumption that the
recheck flag will remain the same from one page to the next. While that
assumption is hardly bulletproof, it seems like a good bet most of the
time, and better than prefetching pages we don't need.
This commit installs the executor infrastructure, but doesn't change
any planner cost estimates, thus possibly causing bitmap scans to
not be chosen in cases where this change renders them the best choice.
I (tgl) am not entirely convinced that we need to account for this
behavior in the planner, because I think typically the bitmap scan would
get chosen anyway if it's the best bet. In any case the submitted patch
took way too many shortcuts, resulting in too many clearly-bad choices,
to be committable.
Alexander Kuzmenkov, reviewed by Alexey Chernyshov, and whacked around
rather heavily by me.
Discussion: https://postgr.es/m/239a8955-c0fc-f506-026d-c837e86c827b@postgrespro.ru
It's possible for dropping a column, or altering its type, to require
changes in domain CHECK constraint expressions; but the code was
previously only expecting to find dependent table CHECK constraints.
Make the necessary adjustments.
This is a fairly old oversight, but it's a lot easier to encounter
the problem in the context of domains over composite types than it
was before. Given the lack of field complaints, I'm not going to
bother with a back-patch, though I'd be willing to reconsider that
decision if someone does complain.
Patch by me, reviewed by Michael Paquier
Discussion: https://postgr.es/m/30656.1509128130@sss.pgh.pa.us
In autovacuum's "work item" processing, a few strings were allocated in
the current transaction's memory context, which goes away during error
handling; if an error happened during execution of the work item, the
pfree() calls to clean up afterwards would try to release already-released
memory, possibly leading to a crash. In branch master, this was already
fixed by commit 335f3d04e4, so backpatch that to REL_10_STABLE to fix
the problem there too.
As a secondary problem, verify that the autovacuum worker is connected
to the right database for each work item; otherwise some items would be
discarded by workers in other databases.
Reported-by: Justin Pryzby
Discussion: https://postgr.es/m/20171014035732.GB31726@telsasoft.com
Without this fix, dropping a role can sometimes result in parallel
query failures in sessions that have used "SET ROLE" to assume the
dropped role, even if that setting isn't active any more.
Report by Pavan Deolasee. Patch by Amit Kapila, reviewed by me.
Discussion: http://postgr.es/m/CABOikdOomRcZsLsLK+Z+qENM1zxyaWnAvFh3MJZzZnnKiF+REg@mail.gmail.com
The previous comment (which was copied as boilerplate from one file
to the next) implied that it was the executor node itself which was
being serialized, but that's not right. We're not serializing
the executor nodes; we're just allowing them to store some
additional information in DSM. Adjusts the comment to reflect this.
Discussion: http://postgr.es/m/CA+TgmoaHVinxG=3h6qBAsyV8xaDyQwbzK7YZnYfE8nJFMK1=FA@mail.gmail.com
Commit d5b760ecb wasn't quite right, on second thought: if the
caller didn't ask for column names then it would happily emit
more Vars than if the caller did ask for column names. This
is surely not a good idea. Advance the aliasp_item whether or
not we're preparing a colnames list.
expandRTE() supposed that an RTE_SUBQUERY subquery must have exactly
as many non-junk tlist items as the RTE has column aliases for it.
This was true at the time the code was written, and is still true so
far as parse analysis is concerned --- but when the function is used
during planning, the subquery might have appeared through insertion
of a view that now has more columns than it did when the outer query
was parsed. This results in a core dump if, for instance, we have
to expand a whole-row Var that references the subquery.
To avoid crashing, we can either stop expanding the RTE when we run
out of aliases, or invent new aliases for the added columns. While
the latter might be more useful, the former is consistent with what
expandRTE() does for composite-returning functions in the RTE_FUNCTION
case, so it seems like we'd better do it that way.
Per bug #14876 from Samuel Horwitz. This has been busted since commit
ff1ea2173 allowed views to acquire more columns, so back-patch to all
supported branches.
Discussion: https://postgr.es/m/20171026184035.1471.82810@wrigleys.postgresql.org
This was always intended to work, but due to an oversight in
max_parallel_hazard_walker, it didn't. In testing, we missed the
fact that it was only working for custom plans, where the parameter
value has been substituted for the parameter itself early enough
that everything worked. In a generic plan, the Param node survives
and must be treated as parallel-safe. SerializeParamList provides
for the transmission of parameter values to workers.
Amit Kapila with help from Kuntal Ghosh. Some changes by me.
Discussion: http://postgr.es/m/CAA4eK1+_BuZrmVCeua5Eqnm4Co9DAXdM5HPAOE2J19ePbR912Q@mail.gmail.com
On closer investigation, commits f3ea3e3e8 et al were a few bricks
shy of a load. What we need is not so much to lock down the result
type of a FieldSelect, as to lock down the existence of the column
it's trying to extract. Otherwise, we can break it by dropping that
column. The dependency on the result type is then held indirectly
through the column, and doesn't need to be recorded explicitly.
Out of paranoia, I left in the code to record a dependency on the
result type, but it's used only if we can't identify the pg_class OID
for the column. That shouldn't ever happen right now, AFAICS, but
it seems possible that in future the input node could be marked as
being of type RECORD rather than some specific composite type.
Likewise for FieldStore.
Like the previous patch, back-patch to all supported branches.
Discussion: https://postgr.es/m/22571.1509064146@sss.pgh.pa.us
If we try to run a parallel plan in serial mode because, for example,
it's going to be scanned via a cursor, but for some reason we're
already in parallel mode (for example because an outer query is
running in parallel), we'd incorrectly try to launch workers.
Fix by adding a flag to the EState, so that we can be certain that
ExecutePlan() and ExecGather()/ExecGatherMerge() will have the same
idea about whether we are executing serially or in parallel.
Report and fix by Amit Kapila with help from Kuntal Ghosh. A few
tweaks by me.
Discussion: http://postgr.es/m/CAA4eK1+_BuZrmVCeua5Eqnm4Co9DAXdM5HPAOE2J19ePbR912Q@mail.gmail.com
This is the last major omission in our domains feature: you can now
make a domain over anything that's not a pseudotype.
The major complication from an implementation standpoint is that places
that might be creating tuples of a domain type now need to be prepared
to apply domain_check(). It seems better that unprepared code fail
with an error like "<type> is not composite" than that it silently fail
to apply domain constraints. Therefore, relevant infrastructure like
get_func_result_type() and lookup_rowtype_tupdesc() has been adjusted
to treat domain-over-composite as a distinct case that unprepared code
won't recognize, rather than just transparently treating it the same
as plain composite. This isn't a 100% solution to the possibility of
overlooked domain checks, but it catches most places.
In passing, improve typcache.c's support for domains (it can now cache
the identity of a domain's base type), and rewrite the argument handling
logic in jsonfuncs.c's populate_record[set]_worker to reduce duplicative
per-call lookups.
I believe this is code-complete so far as the core and contrib code go.
The PLs need varying amounts of work, which will be tackled in followup
patches.
Discussion: https://postgr.es/m/4206.1499798337@sss.pgh.pa.us
Previously, we skipped using search_indexed_tlist_for_sortgroupref()
if the tlist expression being sought in the child plan node was merely
a Var. This is purely an optimization, based on the theory that
search_indexed_tlist_for_var() is faster, and one copy of a Var should
be as good as another. However, the GROUPING SETS patch broke the
latter assumption: grouping columns containing the "same" Var can
sometimes have different outputs, as shown in the test case added here.
So do it the hard way whenever a ressortgroupref marking exists.
(If this seems like a bottleneck, we could imagine building a tlist index
data structure for ressortgroupref values, as we do for Vars. But I'll
let that idea go until there's some evidence it's worthwhile.)
Back-patch to 9.6. The problem also exists in 9.5 where GROUPING SETS
came in, but this patch is insufficient to resolve the problem in 9.5:
there is some obscure dependency on the upper-planner-pathification
work that happened in 9.6. Given that this is such a weird corner case,
and no end users have complained about it, it doesn't seem worth the work
to develop a fix for 9.5.
Patch by me, per a report from Heikki Linnakangas. (This does not fix
Heikki's original complaint, just the follow-on one.)
Discussion: https://postgr.es/m/aefc657e-edb2-64d5-6df1-a0828f6e9104@iki.fi
json_build_object and json_build_array and the jsonb equivalents did not
correctly process explicit VARIADIC arguments. They are modified to use
the new extract_variadic_args() utility function which abstracts away
the details of the call method.
Michael Paquier, reviewed by Tom Lane and Dmitry Dolgov.
Backpatch to 9.5 for the jsonb fixes and 9.4 for the json fixes, as
that's where they originated.
This is epecially useful in the case or "VARIADIC ANY" functions. The
caller can get the artguments and types regardless of whether or not and
explicit VARIADIC array argument has been used. The function also
provides an option to convert arguments on type "unknown" to to "text".
Michael Paquier and me, reviewed by Tom Lane.
Backpatch to 9.4 in order to support the following json bug fix.
Although joinaliasvars lists coming out of the parser are quite simple,
those lists can contain arbitrarily complex expressions after subquery
pullup. We do not perform expression preprocessing on them, meaning that
expressions in those lists will not meet the expectations of later phases
of the planner (for example, that they do not contain SubLinks). This had
been thought pretty harmless, since we don't intentionally touch those
lists in later phases --- but Andreas Seltenreich found a case in which
adjust_appendrel_attrs() could recurse into a joinaliasvars list and then
die on its assertion that it never sees a SubLink. We considered a couple
of localized fixes to prevent that specific case from looking at the
joinaliasvars lists, but really this seems like a generic hazard for all
expression processing in the planner. Therefore, probably the best answer
is to delete the joinaliasvars lists from the parsetree at the end of
expression preprocessing, so that there are no reachable expressions that
haven't been through preprocessing.
The case Andreas found seems to be harmless in non-Assert builds, and so
far there are no field reports suggesting that there are user-visible
effects in other cases. I considered back-patching this anyway, but
it turns out that Andreas' test doesn't fail at all in 9.4-9.6, because
in those versions adjust_appendrel_attrs contains code (added in commit
842faa714 and removed again in commit 215b43cdc) to process SubLinks
rather than complain about them. Barring discovery of another path by
which unprocessed joinaliasvars lists can cause trouble, the most
prudent compromise seems to be to patch this into v10 but not further.
Patch by me, with thanks to Amit Langote for initial investigation
and review.
Discussion: https://postgr.es/m/87r2tvt9f1.fsf@ansel.ydns.eu
find_expr_references() neglected to record a dependency on the result type
of a FieldSelect node, allowing a DROP TYPE to break a view or rule that
contains such an expression. I think we'd omitted this case intentionally,
reasoning that there would always be a related dependency ensuring that the
DROP would cascade to the view. But at least with nested field selection
expressions, that's not true, as shown in bug #14867 from Mansur Galiev.
Add the dependency, and for good measure a dependency on the node's exposed
collation.
Likewise add a dependency on the result type of a FieldStore. I think here
the reasoning was that it'd only appear within an assignment to a field,
and the dependency on the field's column would be enough ... but having
seen this example, I think that's wrong for nested-composites cases.
Looking at nearby code, I notice we're not recording a dependency on the
exposed collation of CoerceViaIO, which seems inconsistent with our choices
for related node types. Maybe that's OK but I'm feeling suspicious of this
code today, so let's add that; it certainly can't hurt.
This patch does not do anything to protect already-existing views, only
views created after it's installed. But seeing that the issue has been
there a very long time and nobody noticed till now, that's probably good
enough.
Back-patch to all supported branches.
Discussion: https://postgr.es/m/20171023150118.1477.19174@wrigleys.postgresql.org
Like the similar logic for arrays and records, it's necessary to examine
the range's subtype to decide whether the range type can support hashing.
We can omit checking the subtype for btree-defined operations, though,
since range subtypes are required to have those operations. (Possibly
that simplification for btree cases led us to overlook that it does
not apply for hash cases.)
This is only an issue if the subtype lacks hash support, which is not
true of any built-in range type, but it's easy to demonstrate a problem
with a range type over, eg, money: you can get a "could not identify
a hash function" failure when the planner is misled into thinking that
hash join or aggregation would work.
This was born broken, so back-patch to all supported branches.
The previous coding would report that an array type supports extended
hashing if its element type supports regular hashing. This bug is
only latent at the moment, since AFAICS there is not yet any code
that depends on checking presence of extended-hashing support to make
any decisions. (And in any case it wouldn't matter unless the element
type has only regular hashing, which isn't true of any core data type.)
But that doesn't make it less broken. Extend the
cache_array_element_properties infrastructure to check this properly.
setTargetTable threw an error if the proposed target RangeVar's relname
matched any visible CTE or ENR. This breaks backwards compatibility in
the CTE case, since pre-v10 we never looked for a CTE here at all, so that
CTE names did not mask regular tables. It does seem like a good idea to
throw an error for the ENR case, though, thus causing ENRs to mask tables
for this purpose; ENRs are new in v10 so we're not breaking existing code,
and we may someday want to allow them to be the targets of DML.
To fix that, replace use of getRTEForSpecialRelationTypes, which was
overkill anyway, with use of scanNameSpaceForENR.
A second problem was that the check neglected to verify null schemaname,
so that a CTE or ENR could incorrectly be thought to match a qualified
RangeVar. That happened because getRTEForSpecialRelationTypes relied
on its caller to have checked for null schemaname. Even though the one
remaining caller got it right, this is obviously bug-prone, so move
the check inside getRTEForSpecialRelationTypes.
Also, revert commit 18ce3a4ab's extremely poorly thought out decision to
add a NULL return case to parserOpenTable --- without either documenting
that or adjusting any of the callers to check for it. The current bug
seems to have arisen in part due to working around that bad idea.
In passing, remove the one-line shim functions transformCTEReference and
transformENRReference --- they don't seem to be adding any clarity or
functionality.
Per report from Hugo Mercier (via Julien Rouhaud). Back-patch to v10
where the bug was introduced.
Thomas Munro, with minor editing by me
Discussion: https://postgr.es/m/CAOBaU_YdPVH+PTtiKSSLOiiW3mVDYsnNUekK+XPbHXiP=wrFLA@mail.gmail.com
Flex generates a lot of functions that are not actually used. In order
to avoid coverage figures being ruined by that, mark up the part of the
.l files where the generated code appears by lcov exclusion markers.
That way, lcov will typically only reported on coverage for the .l file,
which is under our control, but not for the .c file.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
There is no reason to insist that direct arguments must match before
we can merge transition states of two aggregate calls. They're only
used during the finalfn call, so we can treat them as like the finalfn
itself. This allows, eg, merging of
select
percentile_cont(0.25) within group (order by a),
percentile_disc(0.5) within group (order by a)
from ...
This didn't matter (and could not have been tested) before we allowed
state merging of OSAs otherwise.
Discussion: https://postgr.es/m/CAB4ELO5RZhOamuT9Xsf72ozbenDLLXZKSk07FiSVsuJNZB861A@mail.gmail.com
The built-in OSAs all share the same transition function, so they can
share transition state as long as the final functions cooperate to not
do the sort step more than once. To avoid running the tuplesort object
in randomAccess mode unnecessarily, add a bit of infrastructure to
nodeAgg.c to let the aggregate functions find out whether the transition
state is actually being shared or not.
This doesn't work for the hypothetical aggregates, since those inject
a hypothetical row that isn't traceable to the shared input state.
So they remain marked aggfinalmodify = 'w'.
Discussion: https://postgr.es/m/CAB4ELO5RZhOamuT9Xsf72ozbenDLLXZKSk07FiSVsuJNZB861A@mail.gmail.com
An aggregate's input expression(s) are not supposed to be evaluated
at all for a row where its FILTER test fails ... but commit 8ed3f11bb
overlooked that requirement. Reshuffle so that aggregates having a
filter clause evaluate their arguments separately from those without.
This still gets the benefit of doing only one ExecProject in the
common case of multiple Aggrefs, none of which have filters.
While at it, arrange for filter clauses to be included in the common
ExecProject evaluation, thus perhaps buying a little bit even when
there are filters.
Back-patch to v10 where the bug was introduced.
Discussion: https://postgr.es/m/30065.1508161354@sss.pgh.pa.us
While poking around in the aggregate logic, I noticed that commit
8ed3f11bb broke the logic in nodeAgg.c that purports to detect nested
aggregates, by moving initialization of regular aggregate argument
expressions out of the code segment that checks for that.
You could argue that this check is unnecessary, but it's not much code
so I'm inclined to keep it as a backstop against parser and planner
bugs. However, there's certainly zero value in checking only some of
the subexpressions.
We can make the check complete again, and as a bonus make it a good
deal more bulletproof against future mistakes of the same ilk, by
moving it out to the outermost level of ExecInitAgg. This means we
need to check only once per Agg node not once per aggregate, which
also seems like a good thing --- if the check does find something
wrong, it's not urgent that we report it before the plan node
initialization finishes.
Since this requires remembering the original length of the aggs list,
I deleted a long-obsolete stanza that changed numaggs from 0 to 1.
That's so old it predates our decision that palloc(0) is a valid
operation, in (digs...) 2004, see commit 24a1e20f1.
In passing improve a few comments.
Back-patch to v10, just in case.
Up to now, there's been hard-wired assumptions that normal aggregates'
final functions never modify their transition states, while ordered-set
aggregates' final functions always do. This has always been a bit
limiting, and in particular it's getting in the way of improving the
built-in ordered-set aggregates to allow merging of transition states.
Therefore, let's introduce catalog and CREATE AGGREGATE infrastructure
that lets the finalfn's behavior be declared explicitly.
There are now three possibilities for the finalfn behavior: it's purely
read-only, it trashes the transition state irrecoverably, or it changes
the state in such a way that no more transfn calls are possible but the
state can still be passed to other, compatible finalfns. There are no
examples of this third case today, but we'll shortly make the built-in
OSAs act like that.
This change allows user-defined aggregates to explicitly disclaim support
for use as window functions, and/or to prevent transition state merging,
if their implementations cannot handle that. While it was previously
possible to handle the window case with a run-time error check, there was
not any way to prevent transition state merging, which in retrospect is
something commit 804163bc2 should have provided for. But better late
than never.
In passing, split out pg_aggregate.c's extern function declarations into
a new header file pg_aggregate_fn.h, similarly to what we've done for
some other catalog headers, so that pg_aggregate.h itself can be safe
for frontend files to include. This lets pg_dump use the symbolic
names for relevant constants.
Discussion: https://postgr.es/m/4834.1507849699@sss.pgh.pa.us
The following are the individual improvements:
1) Avoidance of FunctionCallInfo based function calls, replaced by
more efficient functions with a native C argument interface.
2) Don't extract columns from a cache entry's tuple whenever matching
entries - instead store them as a Datum array. This also allows to
get rid of having to build dummy tuples for negative & list
entries, and of a hack for dealing with cstring vs. text weirdness.
3) Reorder members of catcache.h struct, so imortant entries are more
likely to be on one cacheline.
4) Allowing the compiler to specialize critical SearchCatCache for a
specific number of attributes allows to unroll loops and avoid
other nkeys dependant initialization.
5) Only initializing the ScanKey when necessary, i.e. catcache misses,
greatly reduces cache unnecessary cpu cache misses.
6) Split of the cache-miss case from the hash lookup, reducing stack
allocations etc in the common case.
7) CatCTup and their corresponding heaptuple are allocated in one
piece.
This results in making cache lookups themselves roughly three times as
fast - full-system benchmarks obviously improve less than that.
I've also evaluated further techniques:
- replace open coded hash with simplehash - the list walk right now
shows up in profiles. Unfortunately it's not easy to do so safely as
an entry's memory location can change at various times, which
doesn't work well with the refcounting and cache invalidation.
- Cacheline-aligning CatCTup entries - helps some with performance,
but the win isn't big and the code for it is ugly, because the
tuples have to be freed as well.
- add more proper functions, rather than macros for
SearchSysCacheCopyN etc., but right now they don't show up in
profiles.
The reason the macro wrapper for syscache.c/h have to be changed,
rather than just catcache, is that doing otherwise would require
exposing the SysCache array to the outside. That might be a good idea
anyway, but it's for another day.
Author: Andres Freund
Reviewed-By: Robert Haas
Discussion: https://postgr.es/m/20170914061207.zxotvyopetm7lrrp@alap3.anarazel.de
If a Parallel Bitmap Heap scan's chain of leftmost descendents
includes a BitmapOr whose first child is a BitmapAnd, the prior coding
would mistakenly create a non-shared TIDBitmap and then try to perform
shared iteration.
Report by Tomas Vondra. Patch by Dilip Kumar.
Discussion: http://postgr.es/m/50e89684-8ad9-dead-8767-c9545bafd3b6@2ndquadrant.com
I (tgl) objected to the obscure implementation introduced in commit
1c497fa72. This one seems a bit less action-at-a-distance-y, at the
price of repeating a few lines of code.
Improve the comments about what the function is doing, too.
Amit Khandekar, whacked around a bit more by me
Discussion: https://postgr.es/m/CAJ3gD9egYTyHUH0nTMxm8-1m3RvdqEbaTyGC-CUNtYf7tKNDaQ@mail.gmail.com
After calling ldap_unbind_s() we probably shouldn't try to use the LDAP
connection again to call ldap_get_option(), even if it failed. The OpenLDAP
man page for ldap_unbind[_s] says "Once it is called, the connection to the
LDAP server is closed, and the ld structure is invalid." Otherwise, as a
general rule we should probably call ldap_unbind() before returning in all
paths to avoid leaking resources. It is unlikely there is any practical
leak problem since failure to authenticate currently results in the backend
exiting soon afterwards.
Author: Thomas Munro
Reviewed-By: Alvaro Herrera, Peter Eisentraut
Discussion: https://postgr.es/m/20170914141205.eup4kxzlkagtmfac%40alvherre.pgsql
The previous convention doesn't lend itself to creating ResultRelInfos
lazily, as we already do in ExecGetTriggerResultRel. This patch
doesn't make anything lazier than before, but the pending patch for
UPDATE tuple routing proposes to do so (and there might be other
opportunities as well).
Amit Khandekar with some adjustments by me.
Discussion: http://postgr.es/m/CA+TgmoYPVP9Lyf6vUFA5DwxS4c--x6LOj2y36BsJaYtp62eXPQ@mail.gmail.com
If we merge the transition calculations for two different aggregates,
it's reasonable to assume that the transition function should not care
which of those Aggref structs it gets from AggGetAggref(). It is not
reasonable to make the same assumption about an aggregate final function,
however. Commit 804163bc2 broke this, as it will pass whichever Aggref
was first associated with the transition state in both cases.
This doesn't create an observable bug so far as the core system is
concerned, because the only existing uses of AggGetAggref() are in
ordered-set aggregates that happen to not pay attention to anything
but the input properties of the Aggref; and besides that, we disabled
sharing of transition calculations for OSAs yesterday. Nonetheless,
if some third-party code were using AggGetAggref() in a normal aggregate,
they would be entitled to call this a bug. Hence, back-patch the fix
to 9.6 where the problem was introduced.
In passing, improve some of the comments about transition state sharing.
Discussion: https://postgr.es/m/CAB4ELO5RZhOamuT9Xsf72ozbenDLLXZKSk07FiSVsuJNZB861A@mail.gmail.com
This ought to work, but the built-in OSAs are not capable of coping,
because their final-functions destructively modify their transition
state (specifically, the contained tuplesort object). That was fine
when those functions were written, but commit 804163bc2 moved the
goalposts without telling orderedsetaggs.c.
We should fix the built-in OSAs to support this, but it will take
a little work, especially if we don't want to sacrifice performance
in the normal non-shared-state case. Given that it took a year after
9.6 release for anyone to notice this bug, we should not prioritize
sharable-state over nonsharable-state performance. And a proper fix
is likely to be more complicated than we'd want to back-patch, too.
Therefore, let's just put in this stop-gap patch to prevent nodeAgg.c
from choosing to use shared state for OSAs. We can revert it in HEAD
when we get a better fix.
Report from Lukas Eder, diagnosis by me, patch by David Rowley.
Back-patch to 9.6 where the problem was introduced.
Discussion: https://postgr.es/m/CAB4ELO5RZhOamuT9Xsf72ozbenDLLXZKSk07FiSVsuJNZB861A@mail.gmail.com
There's three categories of changes leading to better performance:
- Splitting the per-attribute part of SendRowDescriptionMessage into a
v2 and a v3 version allows avoiding branches for every attribute.
- Preallocating the size of the buffer to be big enough for all
attributes and then using pq_write* avoids unnecessary buffer
size checks & resizing.
- Reusing a persistently allocated StringInfo for all
SendRowDescriptionMessage() invocations avoids repeated allocations
& reallocations.
Author: Andres Freund
Discussion: https://postgr.es/m/20170914063418.sckdzgjfrsbekae4@alap3.anarazel.de
This takes advantage of the infrastructure introduced by commit
81c5e46c49 to greatly reduce the
likelihood that two different queries will end up with the same query
ID. It's still possible, of course, but whereas before it the chances
of a collision reached 25% around 50,000 queries, it will now take
more than 3 billion queries.
Backward incompatibility: Because the type exposed at the SQL level is
int8, users may now see negative query IDs in the pg_stat_statements
view (and also, query IDs more than 4 billion, which was the old
limit).
Patch by me, reviewed by Michael Paquier and Peter Geoghegan.
Discussion: http://postgr.es/m/CA+TgmobG_Kp4cBKFmsznUAaM1GWW6hhRNiZC0KjRMOOeYnz5Yw@mail.gmail.com
This avoids newly allocating, and then possibly growing, the
stringbuffer for every row. For wide rows this can substantially
reduce memory allocator overhead, at the price of not immediately
reducing memory usage after outputting an especially wide row.
Author: Andres Freund
Discussion: https://postgr.es/m/20170914063418.sckdzgjfrsbekae4@alap3.anarazel.de
There's three prongs to achieve greater efficiency here:
1) Allow reusing a stringbuffer across pq_beginmessage/endmessage,
with the new pq_beginmessage_reuse/endmessage_reuse. This can be
beneficial both because it avoids allocating the initial buffer,
and because it's more likely to already have an correctly sized
buffer.
2) Replacing pq_sendint() with pq_sendint$width() inline
functions. Previously unnecessary and unpredictable branches in
pq_sendint() were needed. Additionally the replacement functions
are implemented more efficiently. pq_sendint is now deprecated, a
separate commit will convert all in-tree callers.
3) Add pq_writeint$width(), pq_writestring(). These rely on sufficient
space in the StringInfo's buffer, avoiding individual space checks
& potential individual resizing. To allow this to be used for
strings, expose mbutil.c's MAX_CONVERSION_GROWTH.
Followup commits will make use of these facilities.
Author: Andres Freund
Discussion: https://postgr.es/m/20170914063418.sckdzgjfrsbekae4@alap3.anarazel.de
In a lot of the places having appendBinaryStringInfo() maintain a
trailing NUL byte wasn't actually meaningful, e.g. when appending an
integer which can contain 0 in one of its bytes.
Removing this yields some small speedup, but more importantly will be
more consistent when providing faster variants of pq_sendint etc.
Author: Andres Freund
Discussion: https://postgr.es/m/20170914063418.sckdzgjfrsbekae4@alap3.anarazel.de
resowner/README contained advice to use a PG_TRY block to restore the
old CurrentResourceOwner value anywhere that that variable is transiently
changed. That advice was only inconsistently followed, however, and
on reflection it seems like unnecessary overhead. We don't bother
with such a convention for transient CurrentMemoryContext changes,
on the grounds that any (sub)transaction abort will start out by
resetting CurrentMemoryContext to what it wants. But the same is
true of CurrentResourceOwner, so there seems no need to treat it
differently.
Hence, remove PG_TRY blocks that exist only to restore CurrentResourceOwner
before re-throwing the error. There are a couple of places that restore
it along with some other actions, and I left those alone; the restore is
probably unnecessary but no noticeable gain will result from removing it.
Discussion: https://postgr.es/m/5236.1507583529@sss.pgh.pa.us
The previous coding in ProcessInterrupts() could lead to
idle_in_transaction_session_timeout being ignored, when
statement_timeout occurred earlier.
The problem was that ProcessInterrupts() would return before
processing the transaction timeout if QueryCancelPending was set while
QueryCancelHoldoffCount != 0 - which is the case when reading new
commands from the client. Ergo when the idle transaction timeout would
hit.
Fix that by removing the early return. Alternatively the transaction
timeout code could have been moved up, but that early return seems
like an issue that could hit other cases too.
Author: Lukas Fittl
Bug: #14821
Discussion:
https://www.postgresql.org/message-id/20170921010956.17345.61461%40wrigleys.postgresql.orghttps://www.postgresql.org/message-id/CAP53PkxQnv3OWJpyNPGJYT62uY=n1=2CF_Lpc6gVOFnc0-gazw@mail.gmail.com
Backpatch: 9.6-, where idle_in_transaction_session_timeout was introduced.
The GRANT reference page, which lists the default privileges for new
objects, failed to mention that USAGE is granted by default for data
types and domains. As a lesser sin, it also did not specify anything
about the initial privileges for sequences, FDWs, foreign servers,
or large objects. Fix that, and add a comment to acldefault() in the
probably vain hope of getting people to maintain this list in future.
Noted by Laurenz Albe, though I editorialized on the wording a bit.
Back-patch to all supported branches, since they all have this behavior.
Discussion: https://postgr.es/m/1507620895.4152.1.camel@cybertec.at
Up to now async.c has used TransactionIdIsInProgress() to detect whether
a notify message's source transaction is still running. However, that
function has a quick-exit path that reports that XIDs before RecentXmin
are no longer running. If a listening backend is doing nothing but
listening, and not running any queries, there is nothing that will advance
its value of RecentXmin. Once 2 billion transactions elapse, the
RecentXmin check causes active transactions to be reported as not running.
If they aren't committed yet according to CLOG, async.c decides they
aborted and discards their messages. The timing for that is a bit tight
but it can happen when multiple backends are sending notifies concurrently.
The net symptom therefore is that a sufficiently-long-surviving
listen-only backend starts to miss some fraction of NOTIFY traffic,
but only under heavy load.
The only function that updates RecentXmin is GetSnapshotData().
A brute-force fix would therefore be to take a snapshot before
processing incoming notify messages. But that would add cycles,
as well as contention for the ProcArrayLock. We can be smarter:
having taken the snapshot, let's use that to check for running
XIDs, and not call TransactionIdIsInProgress() at all. In this
way we reduce the number of ProcArrayLock acquisitions from one
per message to one per notify interrupt; that's the same under
light load but should be a benefit under heavy load. Light testing
says that this change is a wash performance-wise for normal loads.
I looked around for other callers of TransactionIdIsInProgress()
that might be at similar risk, and didn't find any; all of them
are inside transactions that presumably have already taken a
snapshot.
Problem report and diagnosis by Marko Tiikkaja, patch by me.
Back-patch to all supported branches, since it's been like this
since 9.0.
Discussion: https://postgr.es/m/20170926182935.14128.65278@wrigleys.postgresql.org
The previous placement of the fallback implementation in libpgcommon
was problematic, because libpqport functions need strnlen
functionality.
Move replacement into libpgport. Provide strnlen() under its posix
name, instead of pg_strnlen(). Fix stupid configure bug, executing the
test only when compiled with threading support.
Author: Andres Freund
Discussion: https://postgr.es/m/E1e1gR2-0005fB-SI@gemulon.postgresql.org
Previously nodeProjectSet only released memory once per input tuple,
rather than once per returned tuple. If the computation of an
individual returned tuple requires a lot of memory, that can lead to
problems.
Instead change things so that the expression context can be reset once
per output tuple, which requires a new memory context to store SRF
arguments in.
This is a longstanding issue, but was hard to fix before 9.6, due to
the way tSRFs where evaluated. But it's fairly easy to fix now. We
could backpatch this into 10, but given there've been fewc omplaints
that doesn't seem worth the risk so far.
Reported-By: Lucas Fairchild
Author: Andres Freund, per discussion with Tom Lane
Discussion: https://postgr.es/m/4514.1507318623@sss.pgh.pa.us
copy_file() reads and writes data 64KB at a time (with default BLCKSZ),
and historically has issued a pg_flush_data request after each write.
This turns out to interact really badly with macOS's new APFS file
system: a large file copy takes over 100X longer than it ought to on
APFS, as reported by Brent Dearth. While that's arguably a macOS bug,
it's not clear whether Apple will do anything about it in the near
future, and in any case experimentation suggests that issuing flushes
a bit less often can be helpful on other platforms too.
Hence, rearrange the logic in copy_file() so that flush requests are
issued once per N writes rather than every time through the loop.
I set the FLUSH_DISTANCE to 32MB on macOS (any less than that still
results in a noticeable speed degradation on APFS), but 1MB elsewhere.
In limited testing on Linux and FreeBSD, this seems slightly faster
than the previous code, and certainly no worse. It helps noticeably
on macOS even with the older HFS filesystem.
A simpler change would have been to just increase the size of the
copy buffer without changing the loop logic, but that seems likely
to trash the processor cache without really helping much.
Back-patch to 9.6 where we introduced msync() as an implementation
option for pg_flush_data(). The problem seems specific to APFS's
mmap/msync support, so I don't think we need to go further back.
Discussion: https://postgr.es/m/CADkxhTNv-j2jw2g8H57deMeAbfRgYBoLmVuXkC=YCFBXRuCOww@mail.gmail.com
If the operator is a strict btree equality operator, and X isn't volatile,
then the clause must yield true for any non-null value of X, or null if X
is null. At top level of a WHERE clause, we can ignore the distinction
between false and null results, so it's valid to simplify the clause to
"X IS NOT NULL". This is a useful improvement mainly because we'll get
a far better selectivity estimate in most cases.
Because such cases seldom arise in well-written queries, it is unappetizing
to expend a lot of planner cycles looking for them ... but it turns out
that there's a place we can shoehorn this in practically for free, because
equivclass.c already has to detect and reject candidate equivalences of the
form X = X. That doesn't catch every place that it would be valid to
simplify to X IS NOT NULL, but it catches the typical case. Working harder
doesn't seem justified.
Patch by me, reviewed by Petr Jelinek
Discussion: https://postgr.es/m/CAMjNa7cC4X9YR-vAJS-jSYCajhRDvJQnN7m2sLH1wLh-_Z2bsw@mail.gmail.com
The logical decoding functions do BeginInternalSubTransaction and
RollbackAndReleaseCurrentSubTransaction to clean up after themselves.
It turns out that AtEOSubXact_SPI has an unrecognized assumption that
we always need to cancel the active SPI operation in the SPI context
that surrounds the subtransaction (if there is one). That's true
when the RollbackAndReleaseCurrentSubTransaction call is coming from
the SPI-using function itself, but not when it's happening inside
some unrelated function invoked by a SPI query. In practice the
affected callers are the various PLs.
To fix, record the current subtransaction ID when we begin a SPI
operation, and clean up only if that ID is the subtransaction being
canceled.
Also, remove AtEOSubXact_SPI's assertion that it must have cleaned
up the surrounding SPI context's active tuptable. That's proven
wrong by the same test case.
Also clarify (or, if you prefer, reinterpret) the calling conventions
for _SPI_begin_call and _SPI_end_call. The memory context cleanup
in the latter means that these have always had the flavor of a matched
resource-management pair, but they weren't documented that way before.
Per report from Ben Chobot.
Back-patch to 9.4 where logical decoding came in. In principle,
the SPI changes should go all the way back, since the problem dates
back to commit 7ec1c5a86. But given the lack of field complaints
it seems few people are using internal subtransactions in this way.
So I don't feel a need to take any risks in 9.2/9.3.
Discussion: https://postgr.es/m/73FBA179-C68C-4540-9473-71E865408B15@silentmedia.com
Both ExecMakeFunctionResultSet() and evaluation of simple expressions
need to be done in the per-tuple memory context, not per-query, else
we leak data until end of query. This is a consideration that was
missed while refactoring code in the ProjectSet patch (note that in
pre-v10, ExecMakeFunctionResult is called in the per-tuple context).
Per bug #14843 from Ben M. Diagnosed independently by Andres and myself.
Discussion: https://postgr.es/m/20171005230321.28561.15927@wrigleys.postgresql.org
Sloppy loop coding in set_status_by_pages() resulted in fetching one array
element more than it should from the subxids[] array. The odds of this
resulting in SIGSEGV are pretty small, but we've certainly seen that happen
with similar mistakes elsewhere. While at it, we can get rid of an extra
TransactionIdToPage() calculation per loop.
Per report from David Binderman. Back-patch to all supported branches,
since this code is quite old.
Discussion: https://postgr.es/m/HE1PR0802MB2331CBA919CBFFF0C465EB429C710@HE1PR0802MB2331.eurprd08.prod.outlook.com
When some tuple versions in an update chain are frozen due to them being
older than freeze_min_age, the xmax/xmin trail can become broken. This
breaks HOT (and probably other things). A subsequent VACUUM can break
things in more serious ways, such as leaving orphan heap-only tuples
whose root HOT redirect items were removed. This can be seen because
index creation (or REINDEX) complain like
ERROR: XX000: failed to find parent tuple for heap-only tuple at (0,7) in table "t"
Because of relfrozenxid contraints, we cannot avoid the freezing of the
early tuples, so we must cope with the results: whenever we see an Xmin
of FrozenTransactionId, consider it a match for whatever the previous
Xmax value was.
This problem seems to have appeared in 9.3 with multixact changes,
though strictly speaking it seems unrelated.
Since 9.4 we have commit 37484ad2a "Change the way we mark tuples as
frozen", so the fix is simple: just compare the raw Xmin (still stored
in the tuple header, since freezing merely set an infomask bit) to the
Xmax. But in 9.3 we rewrite the Xmin value to FrozenTransactionId, so
the original value is lost and we have nothing to compare the Xmax with.
To cope with that case we need to compare the Xmin with FrozenXid,
assume it's a match, and hope for the best. Sadly, since you can
pg_upgrade a 9.3 instance containing half-frozen pages to newer
releases, we need to keep the old check in newer versions too, which
seems a bit brittle; I hope we can somehow get rid of that.
I didn't optimize the new function for performance. The new coding is
probably a bit slower than before, since there is a function call rather
than a straight comparison, but I'd rather have it work correctly than
be fast but wrong.
This is a followup after 20b6552242 fixed a few related problems.
Apparently, in 9.6 and up there are more ways to get into trouble, but
in 9.3 - 9.5 I cannot reproduce a problem anymore with this patch, so
there must be a separate bug.
Reported-by: Peter Geoghegan
Diagnosed-by: Peter Geoghegan, Michael Paquier, Daniel Wood,
Yi Wen Wong, Álvaro
Discussion: https://postgr.es/m/CAH2-Wznm4rCrhFAiwKPWTpEw2bXDtgROZK7jWWGucXeH3D1fmA@mail.gmail.com
Instead of joining two partitioned tables in their entirety we can, if
it is an equi-join on the partition keys, join the matching partitions
individually. This involves teaching the planner about "other join"
rels, which are related to regular join rels in the same way that
other member rels are related to baserels. This can use significantly
more CPU time and memory than regular join planning, because there may
now be a set of "other" rels not only for every base relation but also
for every join relation. In most practical cases, this probably
shouldn't be a problem, because (1) it's probably unusual to join many
tables each with many partitions using the partition keys for all
joins and (2) if you do that scenario then you probably have a big
enough machine to handle the increased memory cost of planning and (3)
the resulting plan is highly likely to be better, so what you spend in
planning you'll make up on the execution side. All the same, for now,
turn this feature off by default.
Currently, we can only perform joins between two tables whose
partitioning schemes are absolutely identical. It would be nice to
cope with other scenarios, such as extra partitions on one side or the
other with no match on the other side, but that will have to wait for
a future patch.
Ashutosh Bapat, reviewed and tested by Rajkumar Raghuwanshi, Amit
Langote, Rafia Sabih, Thomas Munro, Dilip Kumar, Antonin Houska, Amit
Khandekar, and by me. A few final adjustments by me.
Discussion: http://postgr.es/m/CAFjFpRfQ8GrQvzp3jA2wnLqrHmaXna-urjm_UY9BqXj=EaDTSA@mail.gmail.com
Discussion: http://postgr.es/m/CAFjFpRcitjfrULr5jfuKWRPsGUX0LQ0k8-yG0Qw2+1LBGNpMdw@mail.gmail.com
If the table attached as a partition is itself partitioned, individual
partitions might have constraints strong enough to skip scanning the
table even if the table actually attached does not. This is pretty
cheap to check, and possibly a big win if it works out.
Amit Langote, with test case changes by me.
Discussion: http://postgr.es/m/1f08b844-0078-aa8d-452e-7af3bf77d05f@lab.ntt.co.jp
Haribabu Kommi, reviewed by Dilip Kumar and Rafia Sabih. Various
cosmetic changes by me to explain why this appears to be safe but
allowing inserts in parallel mode in general wouldn't be. Also, I
removed the REFRESH MATERIALIZED VIEW case from Haribabu's patch,
since I'm not convinced that case is OK, and hacked on the
documentation somewhat.
Discussion: http://postgr.es/m/CAJrrPGdo5bak6qnPWe8Kpi8g_jfQEs-G4SYmG9y+OFaw2-dPvA@mail.gmail.com
Remove obsolete references to get_rel_oids(). Avoid listing specific
relkinds in the comments, since we seem unable to keep such things
in sync with the code, and it's not all that helpful anyhow.
Noted by Michael Paquier, though I rewrote the comments a bit more.
Discussion: https://postgr.es/m/CAB7nPqTWiN9zwKTaOrsnKiGDChqRt7C1+CiiDk4N4OMn92rs6A@mail.gmail.com
A lot of semi-internal code just prints out numeric SPI error codes,
which is not very helpful. We already have an API function to convert
the codes to a string, so let's make more use of that.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
These are two completely unrelated code paths, so it doesn't make sense
to pack them into one function.
Add attribute noreturn to ri_ReportViolation().
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Turns out we have enough functions that the binary search is quite
noticeable in profiles.
Thus have Gen_fmgrtab.pl build a new mapping from a builtin function's
oid to an index in the existing fmgr_builtins array. That keeps the
additional memory usage at a reasonable amount.
Author: Andres Freund, with input from Tom Lane
Discussion: https://postgr.es/m/20170914065128.a5sk7z4xde5uy3ei@alap3.anarazel.de
Not much to say about this; does what it says on the tin.
However, formerly, if there was a column list then the ANALYZE action was
implied; now it must be specified, or you get an error. This is because
it would otherwise be a bit unclear what the user meant if some tables
have column lists and some don't.
Nathan Bossart, reviewed by Michael Paquier and Masahiko Sawada, with some
editorialization by me
Discussion: https://postgr.es/m/E061A8E3-5E3D-494D-94F0-E8A9B312BBFC@amazon.com
Commit 597a87ccc introduced a latch pointer variable to replace use
of a long-lived shared latch in the shared WalRcvData structure.
This was not well thought out, because there are now hazards of the
pointer variable changing while it's being inspected by another
process. This could obviously lead to a core dump in code like
if (WalRcv->latch)
SetLatch(WalRcv->latch);
and there's a more remote risk of a torn read, if we have any
platforms where reading/writing a pointer is not atomic.
An actual problem would occur only if the walreceiver process
exits (gracefully) while the startup process is trying to
signal it, but that seems well within the realm of possibility.
To fix, treat the pointer variable (not the referenced latch)
as being protected by the WalRcv->mutex spinlock. There
remains a race condition that we could apply SetLatch to a
process latch that no longer belongs to the walreceiver, but
I believe that's harmless: at worst it'd cause an extra wakeup
of the next process to use that PGPROC structure.
Back-patch to v10 where the faulty code was added.
Discussion: https://postgr.es/m/22735.1507048202@sss.pgh.pa.us