FreezeXid introduced in a recent commit, so there isn't any data loss in this
approach.
Doing it causes ALTER TABLE (or rather, the forms of it that cause a full table
rewrite) to be affected as well. In this case, the frozen point is RecentXmin,
because after the rewrite all the tuples are relabeled with the rewriting
transaction's Xid.
TOAST tables are fixed automatically as well, as fallout of the way they were
already being handled in the respective code paths.
With this patch, there is no longer need to VACUUM tables for Xid wraparound
purposes that have been cleaned up via TRUNCATE or CLUSTER.
avoid a later needless VACUUM for Xid-wraparound purposes. We can do this
since the table is known to be left empty, so no Xid remains on it.
Per discussion.
there's an indirect dependency on the owner via the parent table. We were
already handling indexes that way, but not toast tables for some reason.
Saves a little catalog space and cuts down the verbosity of checkSharedDependencies
reports.
named foo, would work but the other ordering would not. If a user-specified
type or table name collides with an existing auto-generated array name, just
rename the array type out of the way by prepending more underscores. This
should not create any backward-compatibility issues, since the cases in which
this will happen would have failed outright in prior releases.
Also fix an oversight in the arrays-of-composites patch: ALTER TABLE RENAME
renamed the table's rowtype but not its array type.
needs to check the new constraint against columns of derived domains too.
Also, make it error out if the domain to be modified is used within any
composite-type columns. Eventually we should support that case, but it seems
a bit painful, and not suitable for a back-patch. For the moment just let the
user know we can't do it.
Backpatch to 8.2, which is the only released version that allows nested
domains. Possibly the other part should be back-patched further.
and views (but not system catalogs, nor sequences or toast tables). Get rid
of the hardwired convention that a type's array type is named exactly "_type",
instead using a new column pg_type.typarray to provide the linkage. (It still
will be named "_type", though, except in odd corner cases such as
maximum-length type names.)
Along the way, make tracking of owner and schema dependencies for types more
uniform: a type directly created by the user has these dependencies, while a
table rowtype or auto-generated array type does not have them, but depends on
its parent object instead.
David Fetter, Andrew Dunstan, Tom Lane
rules to be defined with different, per session controllable, behaviors
for replication purposes.
This will allow replication systems like Slony-I and, as has been stated
on pgsql-hackers, other products to control the firing mechanism of
triggers and rewrite rules without modifying the system catalog directly.
The firing mechanisms are controlled by a new superuser-only GUC
variable, session_replication_role, together with a change to
pg_trigger.tgenabled and a new column pg_rewrite.ev_enabled. Both
columns are a single char data type now (tgenabled was a bool before).
The possible values in these attributes are:
'O' - Trigger/Rule fires when session_replication_role is "origin"
(default) or "local". This is the default behavior.
'D' - Trigger/Rule is disabled and fires never
'A' - Trigger/Rule fires always regardless of the setting of
session_replication_role
'R' - Trigger/Rule fires when session_replication_role is "replica"
The GUC variable can only be changed as long as the system does not have
any cached query plans. This will prevent changing the session role and
accidentally executing stored procedures or functions that have plans
cached that expand to the wrong query set due to differences in the rule
firing semantics.
The SQL syntax for changing a triggers/rules firing semantics is
ALTER TABLE <tabname> <when> TRIGGER|RULE <name>;
<when> ::= ENABLE | ENABLE ALWAYS | ENABLE REPLICA | DISABLE
psql's \d command as well as pg_dump are extended in a backward
compatible fashion.
Jan
module and teach PREPARE and protocol-level prepared statements to use it.
In service of this, rearrange utility-statement processing so that parse
analysis does not assume table schemas can't change before execution for
utility statements (necessary because we don't attempt to re-acquire locks
for utility statements when reusing a stored plan). This requires some
refactoring of the ProcessUtility API, but it ends up cleaner anyway,
for instance we can get rid of the QueryContext global.
Still to do: fix up SPI and related code to use the plan cache; I'm tempted to
try to make SQL functions use it too. Also, there are at least some aspects
of system state that we want to ensure remain the same during a replan as in
the original processing; search_path certainly ought to behave that way for
instance, and perhaps there are others.
an opclass for a generic type such as ANYARRAY. The original coding failed
to check that PK and FK columns were of the same array type. Per discussion
with Tom Dunstan. Also, make the code a shade more readable by not trying
to economize on variables.
equality checks it applies, instead of a random dependence on whatever
operators might be named "=". The equality operators will now be selected
from the opfamily of the unique index that the FK constraint depends on to
enforce uniqueness of the referenced columns; therefore they are certain to be
consistent with that index's notion of equality. Among other things this
should fix the problem noted awhile back that pg_dump may fail for foreign-key
constraints on user-defined types when the required operators aren't in the
search path. This also means that the former warning condition about "foreign
key constraint will require costly sequential scans" is gone: if the
comparison condition isn't indexable then we'll reject the constraint
entirely. All per past discussions.
Along the way, make the RI triggers look into pg_constraint for their
information, instead of using pg_trigger.tgargs; and get rid of the always
error-prone fixed-size string buffers in ri_triggers.c in favor of building up
the RI queries in StringInfo buffers.
initdb forced due to columns added to pg_constraint and pg_trigger.
made query plan. Use of ALTER COLUMN TYPE creates a hazard for cached
query plans: they could contain Vars that claim a column has a different
type than it now has. Fix this by checking during plan startup that Vars
at relation scan level match the current relation tuple descriptor. Since
at that point we already have at least AccessShareLock, we can be sure the
column type will not change underneath us later in the query. However,
since a backend's locks do not conflict against itself, there is still a
hole for an attacker to exploit: he could try to execute ALTER COLUMN TYPE
while a query is in progress in the current backend. Seal that hole by
rejecting ALTER TABLE whenever the target relation is already open in
the current backend.
This is a significant security hole: not only can one trivially crash the
backend, but with appropriate misuse of pass-by-reference datatypes it is
possible to read out arbitrary locations in the server process's memory,
which could allow retrieving database content the user should not be able
to see. Our thanks to Jeff Trout for the initial report.
Security: CVE-2007-0556
having md.c return a success/failure boolean to smgr.c, which was just going
to elog anyway, let md.c issue the elog messages itself. This allows better
error reporting, particularly in cases such as "short read" or "short write"
which Peter was complaining of. Also, remove the kluge of allowing mdread()
to return zeroes from a read-beyond-EOF: this is now an error condition
except when InRecovery or zero_damaged_pages = true. (Hash indexes used to
require that behavior, but no more.) Also, enforce that mdwrite() is to be
used for rewriting existing blocks while mdextend() is to be used for
extending the relation EOF. This restriction lets us get rid of the old
ad-hoc defense against creating huge files by an accidental reference to
a bogus block number: we'll only create new segments in mdextend() not
mdwrite() or mdread(). (Again, when InRecovery we allow it anyway, since
we need to allow updates of blocks that were later truncated away.)
Also, clean up the original makeshift patch for bug #2737: move the
responsibility for padding relation segments to full length into md.c.
cases. Operator classes now exist within "operator families". While most
families are equivalent to a single class, related classes can be grouped
into one family to represent the fact that they are semantically compatible.
Cross-type operators are now naturally adjunct parts of a family, without
having to wedge them into a particular opclass as we had done originally.
This commit restructures the catalogs and cleans up enough of the fallout so
that everything still works at least as well as before, but most of the work
needed to actually improve the planner's behavior will come later. Also,
there are not yet CREATE/DROP/ALTER OPERATOR FAMILY commands; the only way
to create a new family right now is to allow CREATE OPERATOR CLASS to make
one by default. I owe some more documentation work, too. But that can all
be done in smaller pieces once this infrastructure is in place.
the target relation(s). There might be some cases where we could discard
the pending event instead, but for the moment a conservative approach
seems sufficient. Per report from Markus Schiltknecht and subsequent
discussion.
by abandoning the idea that it should say SERIAL in the dump. Instead,
dump serial sequences and column defaults just like regular ones.
Add a new backend command ALTER SEQUENCE OWNED BY to let pg_dump recreate
the sequence-to-column dependency that was formerly created "behind the
scenes" by SERIAL. This restores SERIAL to being truly "just a macro"
consisting of component operations that can be stated explicitly in SQL.
Furthermore, the new command allows sequence ownership to be reassigned,
so that old mistakes can be cleaned up.
Also, downgrade the OWNED-BY dependency from INTERNAL to AUTO, since there
is no longer any very compelling argument why the sequence couldn't be
dropped while keeping the column. (This forces initdb, to be sure the
right kinds of dependencies are in there.)
Along the way, add checks to prevent ALTER OWNER or SET SCHEMA on an
owned sequence; you can now only do this indirectly by changing the
owning table's owner or schema. This is an oversight in previous
releases, but probably not worth back-patching.
the DROP pass rather than the ADD_CONSTR pass. On examining the code I
think this was just an oversight rather than intentional, and it seems
to satisfy the principle of least surprise better than the alternative
solution that was discussed. Add an example to the ref page showing how
to do ALTER TYPE and update the default in one command. Per gripe from
Markus Bertheau that that wasn't possible.
(table or index) before trying to open its relcache entry. This fixes
race conditions in which someone else commits a change to the relation's
catalog entries while we are in process of doing relcache load. Problems
of that ilk have been reported sporadically for years, but it was not
really practical to fix until recently --- for instance, the recent
addition of WAL-log support for in-place updates helped.
Along the way, remove pg_am.amconcurrent: all AMs are now expected to support
concurrent update.
created in the bootstrap phase proper, rather than added after-the-fact
by initdb. This is cleaner than before because it allows us to retire the
undocumented ALTER TABLE ... CREATE TOAST TABLE command, but the real reason
I'm doing it is so that toast tables of shared catalogs will now have
predetermined OIDs. This will allow a reasonably clean solution to the
problem of locking tables before we load their relcache entries, to appear
in a forthcoming patch.
a table. Otherwise a USING clause that yields NULL can leave the table
violating its constraint (possibly there are other cases too). Per report
from Alexander Pravking.
discussion (including making def_arg allow reserved words), add missed
opt_definition for UNIQUE case. Put the reloptions support code in a less
random place (I chose to make a new file access/common/reloptions.c).
Eliminate header inclusion creep. Make the index options functions safely
user-callable (seems like client apps might like to be able to test validity
of options before trying to make an index). Reduce overhead for normal case
with no options by allowing rd_options to be NULL. Fix some unmaintainably
klugy code, including getting rid of Natts_pg_class_fixed at long last.
Some stylistic cleanup too, and pay attention to keeping comments in sync
with code.
Documentation still needs work, though I did fix the omissions in
catalogs.sgml and indexam.sgml.
the read lock we hold on the table's parent relation until commit.
Update equalfuncs.c for the new field in AlterTableCmd. Various
improvements to comments, variable names, and error reporting.
There is room for further improvement here, but this is at least
a step in the right direction.
Open items:
There were a few tangentially related issues that have come up that I think
are TODOs. I'm likely to tackle one or two of these next so I'm interested in
hearing feedback on them as well.
. Constraints currently do not know anything about inheritance. Tom suggested
adding a coninhcount and conislocal like attributes have to track their
inheritance status.
. Foreign key constraints currently do not get copied to new children (and
therefore my code doesn't verify them). I don't think it would be hard to
add them and treat them like CHECK constraints.
. No constraints at all are copied to tables defined with LIKE. That makes it
hard to use LIKE to define new partitions. The standard defines LIKE and
specifically says it does not copy constraints. But the standard already has
an option called INCLUDING DEFAULTS; we could always define a non-standard
extension LIKE table INCLUDING CONSTRAINTS that gives the user the option to
request a copy including constraints.
. Personally, I think the whole attislocal thing is bunk. The decision about
whether to drop a column from children tables or not is something that
should be up to the user and trying to DWIM based on whether there was ever
a local definition or the column was acquired purely through inheritance is
hardly ever going to match up with user expectations.
. And of course there's the whole unique and primary key constraint issue. I
think to get any traction at all on this you have a prerequisite of a real
partitioned table implementation where the system knows what the partition
key is so it can recognize when it's a leading part of an index key.
Greg Stark
by creating a reference-count mechanism, similar to what we did a long time
ago for catcache entries. The back branches have an ugly solution involving
lots of extra copies, but this way is more efficient. Reference counting is
only applied to tupdescs that are actually in caches --- there seems no need
to use it for tupdescs that are generated in the executor, since they'll go
away during plan shutdown by virtue of being in the per-query memory context.
Neil Conway and Tom Lane
(relpages/reltuples). To do this, create formal support in heapam.c for
"overwrite" tuple updates (including xlog replay capability) and use that
instead of the ad-hoc overwrites we'd been using in VACUUM and CREATE INDEX.
Take the responsibility for updating stats during CREATE INDEX out of the
individual index AMs, and do it where it belongs, in catalog/index.c. Aside
from being more modular, this avoids having to update the same tuple twice in
some paths through CREATE INDEX. It's probably not measurably faster, but
for sure it's a lot cleaner than before.
during parse analysis, not only errors detected in the flex/bison stages.
This is per my earlier proposal. This commit includes all the basic
infrastructure, but locations are only tracked and reported for errors
involving column references, function calls, and operators. More could
be done later but this seems like a good set to start with. I've also
moved the ReportSyntaxErrorPosition logic out of psql and into libpq,
which should make it available to more people --- even within psql this
is an improvement because warnings weren't handled by ReportSyntaxErrorPosition.
relations are still checked for permissions etc as soon as they are
opened. The original form of the patch could hold exclusive lock for a
long time on relations that the user doesn't even have permissions to
access, let alone truncate.
comment line where output as too long, and update typedefs for /lib
directory. Also fix case where identifiers were used as variable names
in the backend, but as typedefs in ecpg (favor the backend for
indenting).
Backpatch to 8.1.X.
process of dropping roles by dropping objects owned by them and privileges
granted to them, or giving the owned objects to someone else, through the
use of the data stored in the new pg_shdepend catalog.
Some refactoring of the GRANT/REVOKE code was needed, as well as ALTER OWNER
code. Further cleanup of code duplication in the GRANT code seems necessary.
Implemented by me after an idea from Tom Lane, who also provided various kind
of implementation advice.
Regression tests pass. Some tests for the new functionality are also added,
as well as rudimentary documentation.
argument as a 'regclass' value instead of a text string. The frontend
conversion of text string to pg_class OID is now encapsulated as an
implicitly-invocable coercion from text to regclass. This provides
backwards compatibility to the old behavior when the sequence argument
is explicitly typed as 'text'. When the argument is just an unadorned
literal string, it will be taken as 'regclass', which means that the
stored representation will be an OID. This solves longstanding problems
with renaming sequences that are referenced in default expressions, as
well as new-in-8.1 problems with renaming such sequences' schemas or
moving them to another schema. All per recent discussion.
Along the way, fix some rather serious problems in dbmirror's support
for mirroring sequence operations (int4 vs int8 confusion for instance).
the parent table, even if the command that creates them is executed by
someone else (such as a superuser or a member of the owning role).
Per gripe from Michael Fuhr.
use these instead of its previous hack of changing pg_class.reltriggers.
Documentation is lacking, will add that later.
Patch by Satoshi Nagayasu, review and some extra work by Tom Lane.
erroring out as it has done for the last couple weeks. Document that this
form is now ignored because indexes can't usefully have different owners
from their parent tables. Fix pg_dump to not generate ALTER OWNER commands
for indexes.
discussion of getting around this by relaxing the checks made for regular
users, but I'm disinclined to toy with the security model right now,
so just special-case it for superusers where needed.
This was not especially critical before, but it is now that we track
ownership dependencies --- the dependency for the rowtype *must* shift
to the new owner. Spotted by Bernd Helmle.
Also fix a problem introduced by recent change to allow non-superusers
to do ALTER OWNER in some cases: if the table had a toast table, ALTER
OWNER failed *even for superusers*, because the test being applied would
conclude that the new would-be owner had no create rights on pg_toast.
A side-effect of the fix is to disallow changing the ownership of indexes
or toast tables separately from their parent table, which seems a good
idea on the whole.
requiring superuserness always, allow an owner to reassign ownership
to any role he is a member of, if that role would have the right to
create a similar object. These three requirements essentially state
that the would-be alterer has enough privilege to DROP the existing
object and then re-CREATE it as the new role; so we might as well
let him do it in one step. The ALTER TABLESPACE case is a bit
squirrely, but the whole concept of non-superuser tablespace owners
is pretty dubious anyway. Stephen Frost, code review by Tom Lane.
have adequate mechanisms for tracking the contents of databases and
tablespaces). This solves the longstanding problem that you can drop a
user who still owns objects and/or has access permissions.
Alvaro Herrera, with some kibitzing from Tom Lane.
and pg_auth_members. There are still many loose ends to finish in this
patch (no documentation, no regression tests, no pg_dump support for
instance). But I'm going to commit it now anyway so that Alvaro can
make some progress on shared dependencies. The catalog changes should
be pretty much done.
up have the standard layout with unused space between pd_lower and pd_upper.
When this is set, XLogInsert will omit the unused space without bothering
to scan it to see if it's zero. That saves time in XLogInsert, and also
allows reversion of my earlier patch to make PageRepairFragmentation et al
explicitly re-zero freed space. Per suggestion by Heikki Linnakangas.
representation as the jointree) with two lists of RTEs, one showing
the RTEs accessible by qualified names, and the other showing the RTEs
accessible by unqualified names. I think this is conceptually simpler
than what we did before, and it's sure a whole lot easier to search.
This seems to eliminate the parse-time bottleneck for deeply nested
JOIN structures that was exhibited by phil@vodafone.
key, compare the new and old row versions. If the foreign key column has
not changed, we needn't enqueue the trigger, since the update cannot
violate the foreign key. This optimization was previously applied in the
RI trigger function, but it is more efficient to avoid firing the trigger
altogether. Per recent discussion on pgsql-hackers.
Also add a regression test for some unintuitive foreign key behavior, and
refactor some code that deals with the OIDs of the various RI trigger
functions.
keys, rather than a single trigger for both events. This should not change
functionality, but it is more consistent: previously, there were trigger
functions for both "check_insert" and "check_update", but the former was
used for both events.
Bump catalog version number (not strictly necessary, but best to be
cautious).
which is neither needed by nor related to that header. Remove the bogus
inclusion and instead include the header in those C files that actually
need it. Also fix unnecessary inclusions and bad inclusion order in
tsearch2 files.
indexes. Replace all heap_openr and index_openr calls by heap_open
and index_open. Remove runtime lookups of catalog OID numbers in
various places. Remove relcache's support for looking up system
catalogs by name. Bulky but mostly very boring patch ...
indexes. Extend the macros in include/catalog/*.h to carry the info
about hand-assigned OIDs, and adjust the genbki script and bootstrap
code to make the relations actually get those OIDs. Remove the small
number of RelOid_pg_foo macros that we had in favor of a complete
set named like the catname.h and indexing.h macros. Next phase will
get rid of internal use of names for looking up catalogs and indexes;
but this completes the changes forcing an initdb, so it looks like a
good place to commit.
Along the way, I made the shared relations (pg_database etc) not be
'bootstrap' relations any more, so as to reduce the number of hardwired
entries and simplify changing those relations in future. I'm not
sure whether they ever really needed to be handled as bootstrap
relations, but it seems to work fine to not do so now.
of just a relation OID, thereby not having to open the relation for itself.
This actually saves code rather than adding it for most of the existing
callers, which had the rel open already. The main point though is to be
able to use this rather than plain addRangeTableEntry in setTargetTable,
thus saving one relation_openrv/relation_close cycle for every INSERT,
UPDATE, or DELETE. Seems to provide a several percent win on simple
INSERTs.
change saves a great deal of space in pg_proc and its primary index,
and it eliminates the former requirement that INDEX_MAX_KEYS and
FUNC_MAX_ARGS have the same value. INDEX_MAX_KEYS is still embedded
in the on-disk representation (because it affects index tuple header
size), but FUNC_MAX_ARGS is not. I believe it would now be possible
to increase FUNC_MAX_ARGS at little cost, but haven't experimented yet.
There are still a lot of vestigial references to FUNC_MAX_ARGS, which
I will clean up in a separate pass. However, getting rid of it
altogether would require changing the FunctionCallInfoData struct,
and I'm not sure I want to buy into that.
overly strong lock on pg_depend, and it wasn't closing the rel when done.
The latter bug was masked by the ResourceOwner code, which is something
that should be changed.
ExclusiveLock rather than AccessExclusiveLock. This will allow concurrent
SELECT queries to proceed on the table. Per discussion with Andrew at
SuperNews.
to write out data that we are about to tell the filesystem to drop.
smgr_internal_unlink already had a DropRelFileNodeBuffers call to
get rid of dead buffers without a write after it's no longer possible
to roll back the deleting transaction. Adding a similar call in
smgrtruncate simplifies callers and makes the overall division of
labor clearer. This patch removes the former behavior that VACUUM
would write all dirty buffers of a relation unconditionally.
of tuples when passing data up through multiple plan nodes. A slot can now
hold either a normal "physical" HeapTuple, or a "virtual" tuple consisting
of Datum/isnull arrays. Upper plan levels can usually just copy the Datum
arrays, avoiding heap_formtuple() and possible subsequent nocachegetattr()
calls to extract the data again. This work extends Atsushi Ogawa's earlier
patch, which provided the key idea of adding Datum arrays to TupleTableSlots.
(I believe however that something like this was foreseen way back in Berkeley
days --- see the old comment on ExecProject.) A test case involving many
levels of join of fairly wide tables (about 80 columns altogether) showed
about 3x overall speedup, though simple queries will probably not be
helped very much.
I have also duplicated some code in heaptuple.c in order to provide versions
of heap_formtuple and friends that use "bool" arrays to indicate null
attributes, instead of the old convention of "char" arrays containing either
'n' or ' '. This provides a better match to the convention used by
ExecEvalExpr. While I have not made a concerted effort to get rid of uses
of the old routines, I think they should be deprecated and eventually removed.
column with a default expression. In that situation, we need to rewrite
the heap relation. To evaluate the new default expression, we use
ExecEvalExpr(); however, this can allocate memory in the current memory
context, and ATRewriteTable() does not switch out of the active portal's
heap memory context. The end result is a rather large memory leak (on
the order of gigabytes for a reasonably sized table).
This patch changes ATRewriteTable() to switch to the per-tuple memory
context before beginning the per-tuple loop. It also removes an explicit
heap_freetuple() in the loop, since that is no longer needed.
In an unrelated change, I noticed the code was scanning through the
attributes of the new tuple descriptor for each tuple of the old table.
I changed this to use precomputation, which should slightly speed up
the loop.
Thanks to steve@deefs.net for reporting the leak.
command. This is useful because we can allow truncation of tables
referenced by foreign keys, so long as the referencing table is
truncated in the same command.
Alvaro Herrera
is the minimum required fix. I want to look next at taking advantage of
it by simplifying the message semantics in the shared inval message queue,
but that part can be held over for 8.1 if it turns out too ugly.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
more than 65K columns, or when the created table has more than 65K columns
due to adding inherited columns from parent relations. Fix a similar
crash when processing SELECT queries with more than 65K target list
entries. In all three cases we would eventually detect the error and
elog, but the check was being made too late.
clause implicitly whenever one is not given explicitly. Remove concept
of a schema having an associated tablespace, and simplify the rules for
selecting a default tablespace for a table or index. It's now just
(a) explicit TABLESPACE clause; (b) default_tablespace if that's not an
empty string; (c) database's default. This will allow pg_dump to use
SET commands instead of tablespace clauses to determine object locations
(but I didn't actually make it do so). All per recent discussions.
of HeapTupleSatisfiesItself() to trigger a hint-bit update on the tuple:
if the row was updated or deleted by a subtransaction of my own transaction
that was later rolled back. This cannot occur in pre-8.0 of course, so
the hint-bit patch applied a couple weeks ago is OK for existing releases.
But for 8.0 it seems we had better fix things so that RI_FKey_check can
pass the correct buffer number to HeapTupleSatisfiesItself. Accordingly,
add fields to the TriggerData struct to carry the buffer ID(s) for the
old and new tuple(s). There are other possible solutions but this one
seems cleanest; it will allow other AFTER-trigger functions to safely
do tqual.c calls if they want to. Put new fields at end of struct so
that there is no API breakage.
at the top level of the column's old default expression before adding
an implicit coercion to the new column type. This seems to satisfy the
principle of least surprise, as per discussion of bug #1290.
NO ACTION check is deferrable. This seems to be a closer approximation
to what the SQL spec says than what we were doing before, and it prevents
some anomalous behaviors that are possible now that triggers can fire
during the execution of PL functions.
Stephan Szabo.
as per recent discussions. Invent SubTransactionIds that are managed like
CommandIds (ie, counter is reset at start of each top transaction), and
use these instead of TransactionIds to keep track of subtransaction status
in those modules that need it. This means that a subtransaction does not
need an XID unless it actually inserts/modifies rows in the database.
Accordingly, don't assign it an XID nor take a lock on the XID until it
tries to do that. This saves a lot of overhead for subtransactions that
are only used for error recovery (eg plpgsql exceptions). Also, arrange
to release a subtransaction's XID lock as soon as the subtransaction
exits, in both the commit and abort cases. This avoids holding many
unique locks after a long series of subtransactions. The price is some
additional overhead in XactLockTableWait, but that seems acceptable.
Finally, restructure the state machine in xact.c to have a more orthogonal
set of states for subtransactions.
so that we close and flush the doomed relation's relcache entry before
we start to delete the underlying catalog rows, rather than afterwards.
For awhile yesterday I thought that an unexpected relcache entry rebuild
partway through this sequence might explain the infrequent parallel
regression failures we were chasing. It doesn't, mainly because there's
no CommandCounterIncrement in the sequence and so the deletions aren't
"really" done yet. But it sure seems like trouble waiting to happen.
of XLogInsert had the same sort of checkpoint interlock problem as
RecordTransactionCommit, and indeed I found some. Btree index build
and ALTER TABLE SET TABLESPACE write data outside the friendly confines
of the buffer manager, and therefore they have to take their own
responsibility for checkpoint interlock. The easiest solution seems to
be to force smgrimmedsync at the end of the index build or table copy,
even when the operation is being WAL-logged. This is sufficient since
the new index or table will be of interest to no one if we don't get
as far as committing the current transaction.