This fixes two issues with the completion of ALTER TABLE and ALTER INDEX
after SET STATISTICS is typed, trying to suggest schema objects while
the grammar only allows integers.
The tab completion of ALTER INDEX is made smarter by handling properly
more patterns. COLUMN is an optional keyword, but as no column numbers
can be suggested yet as possible input simply adjust the completion so
as no incorrect queries are generated.
Author: Michael Paquier
Reviewed-by: Tatsuro Yamada
Discussion: https://postgr.es/m/20181219092255.GC680@paquier.xyz
At the end of recovery for the post-promotion process, a new history
file is created followed by the last partial segment of the previous
timeline. Based on the timing, the archiver would first try to archive
the last partial segment and then the history file. This can delay the
detection of a new timeline taken, particularly depending on the time it
takes to transfer the last partial segment as it delays the moment the
history file of the new timeline gets archived. This can cause promoted
standbys to use the same timeline as one already taken depending on the
circumstances if multiple instances look at archives at the same
location.
This commit changes the order of archiving so as history files are
archived in priority over other file types, which reduces the likelihood
of the same timeline being taken (still not reducing the window to
zero), and it makes the archiver behave more consistently with the
startup process doing its post-promotion business.
Author: David Steele
Reviewed-by: Michael Paquier, Kyotaro Horiguchi
Discussion: https://postgr.es/m/929068cf-69e1-bba2-9dc0-e05986aed471@pgmasters.net
Backpatch-through: 9.5
COPY can skip writing WAL when loading data on a table which has been
created in the same transaction as the one loading the data, however
this cannot work on views or foreign table as this would result in
trying to flush relation files which do not exist. So disable the
optimization so as commands are able to work the same way with any
configuration of wal_level.
Tests are added to cover the different cases, which need to have
wal_level set to minimal to allow the problem to show up, and that is
not the default configuration.
Reported-by: Luis M. Carril, Etsuro Fujita
Author: Amit Langote, Michael Paquier
Reviewed-by: Etsuro Fujita
Discussion: https://postgr.es/m/15552-c64aa14c5c22f63c@postgresql.org
Backpatch-through: 10, where support for COPY on views has been added,
while v11 has added support for COPY on foreign tables.
In passing, move the list of parameters where it can be used for both
CREATE TABLE and ALTER TABLE, and reorder it alphabetically.
Author: Dagfinn Ilmari Mannsåker
Discussion: https://postgr.es/m/d8j1s77kdbb.fsf@dalvik.ping.uio.no
013ebc0a7b implements so-called GiST microvacuum. That is gistgettuple() marks
index tuples as dead when kill_prior_tuple is set. Later, when new tuple
insertion claims page space, those dead index tuples are physically deleted
from page. When this deletion is replayed on standby, it might conflict with
read-only queries. But 013ebc0a7b doesn't handle this. That may lead to
disappearance of some tuples from read-only snapshots on standby.
This commit implements resolving of conflicts between replay of GiST microvacuum
and standby queries. On the master we implement new WAL record type
XLOG_GIST_DELETE, which comprises necessary information. On stable releases
we've to be tricky to keep WAL compatibility. Information required for conflict
processing is just appended to data of XLOG_GIST_PAGE_UPDATE record. So,
PostgreSQL version, which doesn't know about conflict processing, will just
ignore that.
Reported-by: Andres Freund
Diagnosed-by: Andres Freund
Discussion: https://postgr.es/m/20181212224524.scafnlyjindmrbe6%40alap3.anarazel.de
Author: Alexander Korotkov
Backpatch-through: 9.6
The SQL spec says that sql_identifier is a domain over varchar,
but it also says that that domain is supposed to represent the set
of valid identifiers for the implementation, in particular applying
a length limit matching the implementation's identifier length limit.
We were declaring sql_identifier as just "character varying", thus
duplicating what the spec says about base type, but entirely failing
at the rest of it.
Instead, let's declare sql_identifier as a domain over type "name".
(We can drop the COLLATE "C" added by commit 6b0faf723, since that's
now implicit in "name".) With the recent improvements to name's
comparison support, there's not a lot of functional difference between
name and varchar. So although in principle this is a spec deviation,
it's a pretty minor one. And correctly enforcing PG's name length limit
is a good thing; on balance this seems closer to the intent of the spec
than what we had.
But that's all just language-lawyering. The *real* reason to do this is
that it makes sql_identifier columns exposed by information_schema views
be just direct representations of the underlying "name" catalog columns,
eliminating a semantic mismatch that was disastrous for performance of
typical queries on the information_schema. In combination with the
recent change to allow dropping no-op CoerceToDomain nodes, this allows
(for example) queries such as
select ... from information_schema.tables where table_name = 'foo';
to produce an indexscan rather than a seqscan on pg_class.
Discussion: https://postgr.es/m/CAFj8pRBUCX4LZ2rA2BbEkdD6NN59mgx+BLo1gO08Wod4RLtcTg@mail.gmail.com
information_schema output columns that are declared as being type
sql_identifier are supposed to conform to the implementation's rules
for valid identifiers, in particular the identifier length limit.
Several places potentially violated this limit by concatenating a
function's name and OID. (The OID is added to ensure name uniqueness
within a schema, since the spec doesn't expect function name overloading.)
Simply truncating the concatenation result to fit in "name" won't do,
since losing part of the OID might wind up giving non-unique results.
Instead, let's truncate the function name as necessary.
The most practical way to do that is to do it in a C function; the
information_schema.sql script doesn't have easy access to the value
of NAMEDATALEN, nor does it have an easy way to truncate on the basis
of resulting byte-length rather than number of characters.
(There are still a couple of places that cast concatenation results to
sql_identifier, but as far as I can see they are guaranteed not to produce
over-length strings, at least with the normal value of NAMEDATALEN.)
Discussion: https://postgr.es/m/23817.1545283477@sss.pgh.pa.us
For probably bogus reasons, we acquire only AccessShareLock on the
partition when we try to detach it from its parent partitioned table.
This can cause ugly things to happen if another transaction is doing
any sort of DDL to the partition concurrently.
Upgrade that lock to ShareUpdateExclusiveLock, which per discussion
seems to be the minimum needed.
Reported by Robert Haas.
Discussion: https://postgr.es/m/CA+TgmoYruJQ+2qnFLtF1xQtr71pdwgfxy3Ziy-TxV28M6pEmyA@mail.gmail.com
"$user" in a search_path string is replaced by CURRENT_USER not
SESSION_USER. (It actually was SESSION_USER in the initial implementation,
but we changed it shortly later, and evidently forgot to fix the docs to
match.)
Noted by antonov@stdpr.ru
Discussion: https://postgr.es/m/159151fb45d490c8d31ea9707e9ba99d@stdpr.ru
Using the full width of the CPU's native word size shouldn't cost
anything in typical cases. When working with large bitmapsets,
this halves the number of operations needed for many common BMS
operations. On the right sort of test case, a measurable improvement
is obtained.
David Rowley
Discussion: https://postgr.es/m/CAKJS1f9EGBd2h-VkXvb=51tf+X46zMX5T8h-KYgXEV_u2zmLUw@mail.gmail.com
When a partition is detached from its parent, we acquire locks on all
attached indexes to also detach them ... but we release those locks
immediately. This is a violation of the policy of keeping locks on user
objects to the end of the transaction. Bug introduced in 8b08f7d482.
It's unclear that there are any ill effects possible, but it's clearly
wrong nonetheless. It's likely that bad behavior *is* possible, but
mostly because the relation that the index is for is only locked with
AccessShareLock, which is an older bug that shall be fixed separately.
While touching that line of code, close the index opened with
index_open() using index_close() instead of relation_close().
No difference in practice, but let's be consistent.
Unearthed by Robert Haas.
Discussion: https://postgr.es/m/CA+TgmoYruJQ+2qnFLtF1xQtr71pdwgfxy3Ziy-TxV28M6pEmyA@mail.gmail.com
The following completion patterns are added:
- CREATE TABLE <name> with '(', OF or PARTITION OF.
- CREATE TABLE <name> OF with list of composite types.
- CREATE TABLE name (...) with PARTITION OF, WITH, TABLESPACE, ON
COMMIT (depending on the presence of a temporary table).
- CREATE TABLE ON COMMIT with actions (only for temporary tables).
Author: Dagfinn Ilmari Mannsåker
Discussion: https://postgr.es/m/d8j1s77kdbb.fsf@dalvik.ping.uio.no
The flag for IF NOT EXISTS was only being passed down in the normal
recursing case. It's been this way since originally added in 9.6 in
commit 2cd40adb85 so backpatch back to 9.6.
Now that name comparison has effectively the same behavior as text
comparison, we might as well merge the name_ops opfamily into text_ops,
allowing cross-type comparisons to be processed without forcing a
datatype coercion first. We need do little more than add cross-type
operators to make the opfamily complete, and fix one or two places
in the planner that assumed text_ops was a single-datatype opfamily.
I chose to unify hash name_ops into hash text_ops as well, since the
types have compatible hashing semantics. This allows marking the
new cross-type equality operators as oprcanhash.
(Note: this doesn't remove the name_ops opclasses, so there's no
breakage of index definitions. Those opclasses are just reparented
into the text_ops opfamily.)
Discussion: https://postgr.es/m/15938.1544377821@sss.pgh.pa.us
The "name" comparison operators now all support collations, making them
functionally equivalent to "text" comparisons, except for the different
physical representation of the datatype. They do, in fact, mostly share
the varstr_cmp and varstr_sortsupport infrastructure, which has been
slightly enlarged to handle the case.
To avoid changes in the default behavior of the datatype, set name's
typcollation to C_COLLATION_OID not DEFAULT_COLLATION_OID, so that
by default comparisons to a name value will continue to use strcmp
semantics. (This would have been the case for system catalog columns
anyway, because of commit 6b0faf723, but doing this makes it true for
user-created name columns as well. In particular, this avoids
locale-dependent changes in our regression test results.)
In consequence, tweak a couple of places that made assumptions about
collatable base types always having typcollation DEFAULT_COLLATION_OID.
I have not, however, attempted to relax the restriction that user-
defined collatable types must have that. Hence, "name" doesn't
behave quite like a user-defined type; it acts more like a domain
with COLLATE "C". (Conceivably, if we ever get rid of the need for
catalog name columns to be fixed-length, "name" could actually become
such a domain over text. But that'd be a pretty massive undertaking,
and I'm not volunteering.)
Discussion: https://postgr.es/m/15938.1544377821@sss.pgh.pa.us
Avoid repetitive calls to repalloc() when the required size of the
collector array grows more than 2x in one call. Also ensure that the
array size is a power of 2 (since palloc will probably consume a power
of 2 anyway) and doesn't start out very small (which'd likely just lead
to extra repallocs).
David Rowley, tweaked a bit by me
Discussion: https://postgr.es/m/CAKJS1f8vn-iSBE8PKeVHrnhvyjRNYCxguPFFY08QLYmjWG9hPQ@mail.gmail.com
Remove a comment from the Berkeley days claiming that nbtree must
disambiguate duplicate keys within _bt_moveright(). There is no special
care taken around duplicates within _bt_moveright(), at least since
commit 9e85183bfc removed inscrutable _bt_moveright() code to handle
pages full of duplicates.
Commit 40dae7ec53, which made the handling of interrupted nbtree page
splits more robust, removed an nbtree-specific end-of-recovery cleanup
step. This meant that it was no longer possible to complete an
interrupted page split during recovery. However, a reference to
recovery as a reason for using a NULL stack while inserting into a
parent page was missed. Remove the reference.
Remove a similar obsolete reference to recovery that was introduced much
more recently, as part of the btree fastpath optimization enhancement
that made it into Postgres 11 (commit 2b272734, and follow-up commits).
Backpatch: 11-, where the fastpath optimization was introduced.
Up to now we allowed text columns in system catalogs to use collation
"default", but that isn't really safe because it might mean something
different in template0 than it means in a database cloned from template0.
In particular, this could mean that cloned pg_statistic entries for such
columns weren't entirely valid, possibly leading to bogus planner
estimates, though (probably) not any outright failures.
In the wake of commit 5e0928005, a better solution is available: if we
label such columns with "C" collation, then their pg_statistic entries
will also use that collation and hence will be valid independently of
the database collation.
This also provides a cleaner solution for indexes on such columns than
the hack added by commit 0b28ea79c: the indexes will naturally inherit
"C" collation and don't have to be forced to use text_pattern_ops.
Also, with the planned improvement of type "name" to be collation-aware,
this policy will apply cleanly to both text and name columns.
Because of the pg_statistic angle, we should also apply this policy
to the tables in information_schema. This patch does that by adjusting
information_schema's textual domain types to specify "C" collation.
That has the user-visible effect that order-sensitive comparisons to
textual information_schema view columns will now use "C" collation
by default. The SQL standard says that the collation of those view
columns is implementation-defined, so I think this is legal per spec.
At some point this might allow for translation of such comparisons
into indexable conditions on the underlying "name" columns, although
additional work will be needed before that can happen.
Discussion: https://postgr.es/m/19346.1544895309@sss.pgh.pa.us
"rescanratio" was computed as 1 + rescanned-tuples / total-inner-tuples,
which is sensible if it's to be multiplied by total-inner-tuples or a cost
value corresponding to scanning all the inner tuples. But in reality it
was (mostly) multiplied by inner_rows or a related cost, numbers that take
into account the possibility of stopping short of scanning the whole inner
relation thanks to a limited key range in the outer relation. This'd
still make sense if we could expect that stopping short would result in a
proportional decrease in the number of tuples that have to be rescanned.
It does not, however. The argument that establishes the validity of our
estimate for that number is independent of whether we scan all of the inner
relation or stop short, and experimentation also shows that stopping short
doesn't reduce the number of rescanned tuples. So the correct calculation
is 1 + rescanned-tuples / inner_rows, and we should be sure to multiply
that by inner_rows or a corresponding cost value.
Most of the time this doesn't make much difference, but if we have
both a high rescan rate (due to lots of duplicate values) and an outer
key range much smaller than the inner key range, then the error can
be significant, leading to a large underestimate of the cost associated
with rescanning.
Per report from Vijaykumar Jain. This thinko appears to go all the way
back to the introduction of the rescan estimation logic in commit
70fba7043, so back-patch to all supported branches.
Discussion: https://postgr.es/m/CAE7uO5hMb_TZYJcZmLAgO6iD68AkEK6qCe7i=vZUkCpoKns+EQ@mail.gmail.com
The new grammar pattern of ALTER INDEX SET STATISTICS able to use column
numbers on top of the existing column names introduced by commit 5b6d13e
forgot to add support for the feature in pg_dump, so defining statistics
on index columns was missing from the dumps, potentially causing silent
planning problems with a subsequent restore.
pg_dump ought to not use column names in what it generates as these are
automatically generated by the server and could conflict with real
relation attributes with matching patterns. "expr" and "exprN", N
incremented automatically after the creation of the first one, are used
as default attribute names for index expressions, and that could easily
match what is defined in other relations, causing the dumps to fail if
some of those attributes are renamed at some point. So to avoid any
problems, the new grammar with column numbers gets used.
Reported-by: Ronan Dunklau
Author: Michael Paquier
Reviewed-by: Tom Lane, Adrien Nayrat, Amul Sul
Discussion: https://postgr.es/m/CAARsnT3UQ4V=yDNW468w8RqHfYiY9mpn2r_c5UkBJ97NAApUEw@mail.gmail.com
Backpatch-through: 11, where the new syntax has been introduced.
It appears that all platforms that have sys_siglist[] also have
strsignal(), making that fallback case in pg_strsignal() dead code.
Getting rid of it allows dropping a configure test, which seems worth
more than providing textual signal descriptions on whatever platforms
might still hypothetically have use for the fallback case.
Discussion: https://postgr.es/m/25758.1544983503@sss.pgh.pa.us
When partitioned tables were introduced, we failed to realize that by
copying the tablespace handling for other relation kinds with no
physical storage we were causing the secondary effect that their
partitions would not automatically inherit the tablespace setting. This
is surprising and unhelpful, so change it to adopt the behavior
introduced in pg11 (commit 33e6c34c32) for partitioned indexes: the
parent relation remembers the tablespace specification, which is then
used for any new partitions that don't declare one.
Because this commit changes behavior of the TABLESPACE clause for
partitioned tables (it's no longer a no-op), it is not backpatched.
Author: David Rowley, Álvaro Herrera
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/CAKJS1f9SxVzqDrGD1teosFd6jBMM0UEaa14_8mRvcWE19Tu0hA@mail.gmail.com
This is an oversight from recent commit b13fd344. While on it, tweak
the previous test with a better name for the renamed primary key.
Detected by buildfarm member prion which forces relation cache release
with -DRELCACHE_FORCE_RELEASE. Back-patch down to 9.4 as the previous
commit.
When a constraint gets renamed, it may have associated with it a target
relation (for example domain constraints don't have one). Not
invalidating the target relation cache when issuing the renaming can
result in issues with subsequent commands that refer to the old
constraint name using the relation cache, causing various failures. One
pattern spotted was using CREATE TABLE LIKE after a constraint
renaming.
Reported-by: Stuart <sfbarbee@gmail.com>
Author: Amit Langote
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/2047094.V130LYfLq4@station53.ousa.org
At least as far back as the 2008 spec, POSIX has defined strsignal(3)
for looking up descriptive strings for signal numbers. We hadn't gotten
the word though, and were still using the crufty old sys_siglist array,
which is in no standard even though most Unixen provide it.
Aside from not being formally standards-compliant, this was just plain
ugly because it involved #ifdef's at every place using the code.
To eliminate the #ifdef's, create a portability function pg_strsignal,
which wraps strsignal(3) if available and otherwise falls back to
sys_siglist[] if available. The set of Unixen with neither API is
probably empty these days, but on any platform with neither, you'll
just get "unrecognized signal". All extant callers print the numeric
signal number too, so no need to work harder than that.
Along the way, upgrade pg_basebackup's child-error-exit reporting
to match the rest of the system.
Discussion: https://postgr.es/m/25758.1544983503@sss.pgh.pa.us
reap_child() basically ignored the possibility of either an error in
waitpid() itself or a child process failure on signal. We don't really
need to do more than report and crash hard, but proceeding as though
nothing is wrong is definitely Not Acceptable. The error report for
nonzero child exit status was pretty off-point, as well.
Noted while fooling around with child-process failure detection
logic elsewhere. It's been like this a long time, so back-patch to
all supported branches.
Commit ffa4cbd62 added logic to detect SIGPIPE failure of a COPY child
process, but it only worked correctly if the SIGPIPE occurred in the
immediate child process. Depending on the shell in use and the
complexity of the shell command string, we might instead get back
an exit code of 128 + SIGPIPE, representing a shell error exit
reporting SIGPIPE in the child process.
We could just hack up ClosePipeToProgram() to add the extra case,
but it seems like this is a fairly general issue deserving a more
general and better-documented solution. I chose to add a couple
of functions in src/common/wait_error.c, which is a natural place
to know about wait-result encodings, that will test for either a
specific child-process signal type or any child-process signal failure.
Then, adjust other places that were doing ad-hoc tests of this type
to use the common functions.
In RestoreArchivedFile, this fixes a race condition affecting whether
the process will report an error or just silently proc_exit(1): before,
that depended on whether the intermediate shell got SIGTERM'd itself
or reported a child process failing on SIGTERM.
Like the previous patch, back-patch to v10; we could go further
but there seems no real need to.
Per report from Erik Rijkers.
Discussion: https://postgr.es/m/f3683f87ab1701bea5d86a7742b22432@xs4all.nl
When we first put in collations support, we basically punted on teaching
pg_statistic, ANALYZE, and the planner selectivity functions about that.
They've just used DEFAULT_COLLATION_OID independently of the actual
collation of the data. It's time to improve that, so:
* Add columns to pg_statistic that record the specific collation associated
with each statistics slot.
* Teach ANALYZE to use the column's actual collation when comparing values
for statistical purposes, and record this in the appropriate slot. (Note
that type-specific typanalyze functions are now expected to fill
stats->stacoll with the appropriate collation, too.)
* Teach assorted selectivity functions to use the actual collation of
the stats they are looking at, instead of just assuming it's
DEFAULT_COLLATION_OID.
This should give noticeably better results in selectivity estimates for
columns with nondefault collations, at least for query clauses that use
that same collation (which would be the default behavior in most cases).
It's still true that comparisons with explicit COLLATE clauses different
from the stored data's collation won't be well-estimated, but that's no
worse than before. Also, this patch does make the first step towards
doing better with that, which is that it's now theoretically possible to
collect stats for a collation other than the column's own collation.
Patch by me; thanks to Peter Eisentraut for review.
Discussion: https://postgr.es/m/14706.1544630227@sss.pgh.pa.us
The cache lookup routines for foreign-data wrappers and foreign servers
are extended with an extra argument to handle a set of flags. The only
value which can be used now is to indicate if a missing object should
result in an error or not, and are designed to be extensible on need.
Those new routines are added into the existing set of user-visible
FDW APIs and documented in consequence. They will be used for future
patches to improve the SQL interface for object addresses.
Author: Michael Paquier
Reviewed-by: Álvaro Herrera
Discussion: https://postgr.es/m/CAB7nPqSZxrSmdHK-rny7z8mi=EAFXJ5J-0RbzDw6aus=wB5azQ@mail.gmail.com
The changes I made in 578b229718 assigned oids below
FirstBootstrapObjectId to objects in include/catalog/*.dat files that
did not have an oid assigned, starting at the max oid explicitly
assigned. Tom criticized that for mainly two reasons:
1) It's not clear which values are manually and which explicitly
assigned.
2) The space below FirstBootstrapObjectId gets pretty crowded, and
some PostgreSQL forks have used oids >= 9000 for their own objects,
to avoid conflicting.
Thus create a new range for objects not assigned explicit oids, but
assigned by genbki.pl. For now 1-9999 is for explicitly assigned oids,
FirstGenbkiObjectId (10000) to FirstBootstrapObjectId (1200) -1 is for
genbki.pl assigned oids, and < FirstNormalObjectId (16384) is for oids
assigned during bootstrap. It's possible that we'll have to adjust
these boundaries, but there's some headroom for now.
Add a note suggesting that oids in forks should be assigned in the
9000-9999 range.
Catversion bump for obvious reasons.
Per complaint from Tom Lane.
Author: Andres Freund
Discussion: https://postgr.es/m/16845.1544393682@sss.pgh.pa.us
The idea here is to not call recordDependencyOn for the default collation,
since we know that's pinned. But what the code actually did was to record
the partition key's dependency on the opclass twice, instead.
Evidently introduced by sloppy coding in commit 2186b608b. Back-patch
to v10 where that came in.
If a domain has no constraints, then CoerceToDomain doesn't really do
anything and can be simplified to a RelabelType. This not only
eliminates cycles at execution, but allows the planner to optimize better
(for instance, match the coerced expression to an index on the underlying
column). However, we do have to support invalidating the plan later if
a constraint gets added to the domain. That's comparable to the case of
a change to a SQL function that had been inlined into a plan, so all the
necessary logic already exists for plans depending on functions. We
need only duplicate or share that logic for domains.
ALTER DOMAIN ADD/DROP CONSTRAINT need to be taught to send out sinval
messages for the domain's pg_type entry, since those operations don't
update that row. (ALTER DOMAIN SET/DROP NOT NULL do update that row,
so no code change is needed for them.)
Testing this revealed what's really a pre-existing bug in plpgsql:
it caches the SQL-expression-tree expansion of type coercions and
had no provision for invalidating entries in that cache. Up to now
that was only a problem if such an expression had inlined a SQL
function that got changed, which is unlikely though not impossible.
But failing to track changes of domain constraints breaks an existing
regression test case and would likely cause practical problems too.
We could fix that locally in plpgsql, but what seems like a better
idea is to build some generic infrastructure in plancache.c to store
standalone expressions and track invalidation events for them.
(It's tempting to wonder whether plpgsql's "simple expression" stuff
could use this code with lower overhead than its current use of the
heavyweight plancache APIs. But I've left that idea for later.)
Other stuff fixed in passing:
* Allow estimate_expression_value() to drop CoerceToDomain
unconditionally, effectively assuming that the coercion will succeed.
This will improve planner selectivity estimates for cases involving
estimatable expressions that are coerced to domains. We could have
done this independently of everything else here, but there wasn't
previously any need for eval_const_expressions_mutator to know about
CoerceToDomain at all.
* Use a dlist for plancache.c's list of cached plans, rather than a
manually threaded singly-linked list. That eliminates a potential
performance problem in DropCachedPlan.
* Fix a couple of inconsistencies in typecmds.c about whether
operations on domains drop RowExclusiveLock on pg_type. Our common
practice is that DDL operations do drop catalog locks, so standardize
on that choice.
Discussion: https://postgr.es/m/19958.1544122124@sss.pgh.pa.us
When GIN vacuum deletes a posting tree page, it assumes that no concurrent
searchers can access it, thanks to ginStepRight() locking two pages at once.
However, since 9.4 searches can skip parts of posting trees descending from the
root. That leads to the risk that page is deleted and reclaimed before
concurrent search can access it.
This commit prevents the risk of above by waiting for every transaction, which
might wait to reference this page, to finish. Due to binary compatibility
we can't change GinPageOpaqueData to store corresponding transaction id.
Instead we reuse page header pd_prune_xid field, which is unused in index pages.
Discussion: https://postgr.es/m/31a702a.14dd.166c1366ac1.Coremail.chjischj%40163.com
Author: Andrey Borodin, Alexander Korotkov
Reviewed-by: Alexander Korotkov
Backpatch-through: 9.4
On standby ginRedoDeletePage() can work concurrently with read-only queries.
Those queries can traverse posting tree in two ways.
1) Using rightlinks by ginStepRight(), which locks the next page before
unlocking its left sibling.
2) Using downlinks by ginFindLeafPage(), which locks at most one page at time.
Original lock order was: page, parent, left sibling. That lock order can
deadlock with ginStepRight(). In order to prevent deadlock this commit changes
lock order to: left sibling, page, parent. Note, that position of parent in
locking order seems insignificant, because we only lock one page at time while
traversing downlinks.
Reported-by: Chen Huajun
Diagnosed-by: Chen Huajun, Peter Geoghegan, Andrey Borodin
Discussion: https://postgr.es/m/31a702a.14dd.166c1366ac1.Coremail.chjischj%40163.com
Author: Alexander Korotkov
Backpatch-through: 9.4
Before 218f51584d if posting tree page is about to be deleted, then the whole
posting tree is locked by LockBufferForCleanup() on root preventing all the
concurrent inserts. 218f51584d reduced locking to the subtree containing
page to be deleted. However, due to concurrent parent split, inserter doesn't
always holds pins on all the pages constituting path from root to the target
leaf page. That could cause a deadlock between GIN vacuum process and GIN
inserter. And we didn't find non-invasive way to fix this.
This commit reverts VACUUM behavior to lock the whole posting tree before
delete any page. However, we keep another useful change by 218f51584d5: the
tree is locked only if there are pages to be deleted.
Reported-by: Chen Huajun
Diagnosed-by: Chen Huajun, Andrey Borodin, Peter Geoghegan
Discussion: https://postgr.es/m/31a702a.14dd.166c1366ac1.Coremail.chjischj%40163.com
Author: Alexander Korotkov, based on ideas from Andrey Borodin and Peter Geoghegan
Reviewed-by: Andrey Borodin
Backpatch-through: 10
postgres_fdw's postgresGetForeignPlan() assumes without checking that the
outer_plan it's given for a join relation must have a NestLoop, MergeJoin,
or HashJoin node at the top. That's been wrong at least since commit
4bbf6edfb (which could cause insertion of a Sort node on top) and it seems
like a pretty unsafe thing to Just Assume even without that.
Through blind good fortune, this doesn't seem to have any worse
consequences today than strange EXPLAIN output, but it's clearly trouble
waiting to happen.
To fix, test the node type explicitly before touching Join-specific
fields, and avoid jamming the new tlist into a node type that can't
do projection. Export a new support function from createplan.c
to avoid building low-level knowledge about the latter into FDWs.
Back-patch to 9.6 where the faulty coding was added. Note that the
associated regression test cases don't show any changes before v11,
apparently because the tests back-patched with 4bbf6edfb don't actually
exercise the problem case before then (there's no top-level Sort
in those plans).
Discussion: https://postgr.es/m/8946.1544644803@sss.pgh.pa.us
Our support for multiple-set-clauses in UPDATE assumes that the Params
referencing a MULTIEXPR_SUBLINK SubPlan will appear before that SubPlan
in the targetlist of the plan node that calculates the updated row.
(Yeah, it's a hack...) In some PG branches it's possible that a Result
node gets inserted between the primary calculation of the update tlist
and the ModifyTable node. setrefs.c did the wrong thing in this case
and left the upper-level Params as Params, causing a crash at runtime.
What it should do is replace them with "outer" Vars referencing the child
plan node's output. That's a result of careless ordering of operations
in fix_upper_expr_mutator, so we can fix it just by reordering the code.
Fix fix_join_expr_mutator similarly for consistency, even though join
nodes could never appear in such a context. (In general, it seems
likely to be a bit cheaper to use Vars than Params in such situations
anyway, so this patch might offer a tiny performance improvement.)
The hazard extends back to 9.5 where the MULTIEXPR_SUBLINK stuff
was introduced, so back-patch that far. However, this may be a live
bug only in 9.6.x and 10.x, as the other branches don't seem to want
to calculate the final tlist below the Result node. (That plan shape
change between branches might be a mini-bug in itself, but I'm not
really interested in digging into the reasons for that right now.
Still, add a regression test memorializing what we expect there,
so we'll notice if it changes again.)
Per bug report from Eduards Bezverhijs.
Discussion: https://postgr.es/m/b6cd572a-3e44-8785-75e9-c512a5a17a73@tieto.com
Michael Paquier