If a tuple in a syscache contains an out-of-line toasted field, and we
try to fetch that field shortly after some other transaction has committed
an update or deletion of the tuple, there is a race condition: vacuum
could come along and remove the toast tuples before we can fetch them.
This leads to transient failures like "missing chunk number 0 for toast
value NNNNN in pg_toast_2619", as seen in recent reports from Andrew
Hammond and Tim Uckun.
The design idea of syscache is that access to stale syscache entries
should be prevented by relation-level locks, but that fails for at least
two cases where toasted fields are possible: ANALYZE updates pg_statistic
rows without locking out sessions that might want to plan queries on the
same table, and CREATE OR REPLACE FUNCTION updates pg_proc rows without
any meaningful lock at all.
The least risky fix seems to be an idea that Heikki suggested when we
were dealing with a related problem back in August: forcibly detoast any
out-of-line fields before putting a tuple into syscache in the first place.
This avoids the problem because at the time we fetch the parent tuple from
the catalog, we should be holding an MVCC snapshot that will prevent
removal of the toast tuples, even if the parent tuple is outdated
immediately after we fetch it. (Note: I'm not convinced that this
statement holds true at every instant where we could be fetching a syscache
entry at all, but it does appear to hold true at the times where we could
fetch an entry that could have a toasted field. We will need to be a bit
wary of adding toast tables to low-level catalogs that don't have them
already.) An additional benefit is that subsequent uses of the syscache
entry should be faster, since they won't have to detoast the field.
Back-patch to all supported versions. The problem is significantly harder
to reproduce in pre-9.0 releases, because of their willingness to flush
every entry in a syscache whenever the underlying catalog is vacuumed
(cf CatalogCacheFlushRelation); but there is still a window for trouble.
Add a column pg_class.relallvisible to remember the number of pages that
were all-visible according to the visibility map as of the last VACUUM
(or ANALYZE, or some other operations that update pg_class.relpages).
Use relallvisible/relpages, instead of an arbitrary constant, to estimate
how many heap page fetches can be avoided during an index-only scan.
This is pretty primitive and will no doubt see refinements once we've
acquired more field experience with the index-only scan mechanism, but
it's way better than using a constant.
Note: I had to adjust an underspecified query in the window.sql regression
test, because it was changing answers when the plan changed to use an
index-only scan. Some of the adjacent tests perhaps should be adjusted
as well, but I didn't do that here.
When a btree index contains all columns required by the query, and the
visibility map shows that all tuples on a target heap page are
visible-to-all, we don't need to fetch that heap page. This patch depends
on the previous patches that made the visibility map reliable.
There's a fair amount left to do here, notably trying to figure out a less
chintzy way of estimating the cost of an index-only scan, but the core
functionality seems ready to commit.
Robert Haas and Ibrar Ahmed, with some previous work by Heikki Linnakangas.
The code path that tried a generic plan, didn't like it, and then made a
custom plan was mistakenly passing the same copy of the query_list to the
planner both times. This doesn't work too well for nontrivial queries,
since the planner tends to scribble on its input. Diagnosis and fix by
Yamamoto Takashi.
I had copied-and-pasted a claim that we couldn't reach this point when
dealing with utility statements, but that was a leftover from when the
caller was required to supply a plan to start with. We now will go
through here at least once when handling a utility statement, so it
seems worth a check to see whether a snapshot is actually needed.
(Note that analyze_requires_snapshot is quite a cheap test.)
Per suggestion from Yamamoto Takashi. I don't think I believe that this
resolves his reported assertion failure; but it's worth changing anyway,
just to save a cycle or two.
The regression tests were failing with CLOBBER_CACHE_ALWAYS enabled,
as reported by buildfarm member jaguar. There was an Assert in
BuildCachedPlan that asserted that the CachedPlanSource hadn't been
invalidated since we called RevalidateCachedQuery, which in theory can't
happen because we are holding locks on all the relevant database objects.
However, CLOBBER_CACHE_ALWAYS generates a false positive by making an
invalidation happen anyway; and on reflection, that could also occur as a
result of a badly-timed sinval reset due to queue overflow. We could just
remove the Assert and forge ahead with the not-really-stale querytree, but
it seems safer to do another RevalidateCachedQuery call just to make real
sure everything's OK.
Rewrite plancache.c so that a "cached plan" (which is rather a misnomer
at this point) can support generation of custom, parameter-value-dependent
plans, and can make an intelligent choice between using custom plans and
the traditional generic-plan approach. The specific choice algorithm
implemented here can probably be improved in future, but this commit is
all about getting the mechanism in place, not the policy.
In addition, restructure the API to greatly reduce the amount of extraneous
data copying needed. The main compromise needed to make that possible was
to split the initial creation of a CachedPlanSource into two steps. It's
worth noting in particular that SPI_saveplan is now deprecated in favor of
SPI_keepplan, which accomplishes the same end result with zero data
copying, and no need to then spend even more cycles throwing away the
original SPIPlan. The risk of long-term memory leaks while manipulating
SPIPlans has also been greatly reduced. Most of this improvement is based
on use of the recently-added MemoryContextSetParent primitive.
walsender.h should depend on xlog.h, not vice versa. (Actually, the
inclusion was circular until a couple hours ago, which was even sillier;
but Bruce broke it in the expedient rather than logically correct
direction.) Because of that poor decision, plus blind application of
pgrminclude, we had a situation where half the system was depending on
xlog.h to include such unrelated stuff as array.h and guc.h. Clean up
the header inclusion, and manually revert a lot of what pgrminclude had
done so things build again.
This episode reinforces my feeling that pgrminclude should not be run
without adult supervision. Inclusion changes in header files in particular
need to be reviewed with great care. More generally, it'd be good if we
had a clearer notion of module layering to dictate which headers can sanely
include which others ... but that's a big task for another day.
It is possible for VACUUM to scan no pages at all, if the visibility map
shows that all pages are all-visible. In this situation VACUUM has no new
information to report about the relation's tuple density, so it wasn't
changing pg_class.reltuples ... but it updated pg_class.relpages anyway.
That's wrong in general, since there is no evidence to justify changing the
density ratio reltuples/relpages, but it's particularly bad if the previous
state was relpages=reltuples=0, which means "unknown tuple density".
We just replaced "unknown" with "zero". ANALYZE would eventually recover
from this, but it could take a lot of repetitions of ANALYZE to do so if
the relation size is much larger than the maximum number of pages ANALYZE
will scan, because of the moving-average behavior introduced by commit
b4b6923e03.
The only known situation where we could have relpages=reltuples=0 and yet
the visibility map asserts everything's visible is immediately following
a pg_upgrade. It might be advisable for pg_upgrade to try to preserve the
relpages/reltuples statistics; but in any case this code is wrong on its
own terms, so fix it. Per report from Sergey Koposov.
Back-patch to 8.4, where the visibility map was introduced, same as the
previous change.
This requires adjusting the API for syscache callback functions: they now
get a hash value, not a TID, to identify the target tuple. Most of them
weren't paying any attention to that argument anyway, but plancache did
require a small amount of fixing.
Also, improve performance a trifle by avoiding sending duplicate inval
messages when a heap_update isn't changing the catcache lookup columns.
The TID isn't stable enough: we might queue an sinval event before a VACUUM
FULL, and then process it afterwards, when the target tuple no longer has
the same TID. So we must invalidate entries on the basis of hash value
only. The old coding can be shown to result in various bizarre,
hard-to-reproduce errors in the presence of concurrent VACUUM FULLs on
system catalogs, and could easily result in permanent catalog corruption,
up to and including complete loss of tables.
This commit is just a minimal fix that removes the unsafe comparison.
We should remove transmission of the tuple TID from sinval messages
altogether, and then arrange to suppress the extra message in the common
case of a heap_update that doesn't change the key hashvalue. But that's
going to be much more invasive, and will only produce a probably-marginal
performance gain, so it doesn't seem like material for a back-patch.
Back-patch to 9.0. Before that, VACUUM FULL refused to do any tuple moving
if it found any INSERT_IN_PROGRESS or DELETE_IN_PROGRESS tuples (and
CLUSTER would give up altogether), so there was no risk of moving a tuple
that might be the subject of an unsent sinval message.
We have to be sure that we have revalidated each nailed-in-cache relcache
entry before we try to use it to load data for some other relcache entry.
The introduction of "mapped relations" in 9.0 broke this, because although
we updated the state kept in relmapper.c early enough, we failed to
propagate that information into relcache entries soon enough; in
particular, we could try to fetch pg_class rows out of pg_class before
we'd updated its relcache entry's rd_node.relNode value from the map.
This bug accounts for Dave Gould's report of failures after "vacuum full
pg_class", and I believe that there is risk for other system catalogs
as well.
The core part of the fix is to copy relmapper data into the relcache
entries during "phase 1" in RelationCacheInvalidate(), before they'll be
used in "phase 2". To try to future-proof the code against other similar
bugs, I also rearranged the order in which nailed relations are visited
during phase 2: now it's pg_class first, then pg_class_oid_index, then
other nailed relations. This should ensure that RelationClearRelation can
apply RelationReloadIndexInfo to all nailed indexes without risking use
of not-yet-revalidated relcache entries.
Back-patch to 9.0 where the relation mapper was introduced.
The previous code tried to synchronize by unlinking the init file twice,
but that doesn't actually work: it leaves a window wherein a third process
could read the already-stale init file but miss the SI messages that would
tell it the data is stale. The result would be bizarre failures in catalog
accesses, typically "could not read block 0 in file ..." later during
startup.
Instead, hold RelCacheInitLock across both the unlink and the sending of
the SI messages. This is more straightforward, and might even be a bit
faster since only one unlink call is needed.
This has been wrong since it was put in (in 2002!), so back-patch to all
supported releases.
get_op_btree_interpretation assumed this in order to save some duplication
of code, but it's not true in general anymore because we added <> support
to btree_gist. (We still assume it for btree opclasses, though.)
Also, essentially the same logic was baked into predtest.c. Get rid of
that duplication by generalizing get_op_btree_interpretation so that it
can be used by predtest.c.
Per bug report from Denis de Bernardy and investigation by Jeff Davis,
though I didn't use Jeff's patch exactly as-is.
Back-patch to 9.1; we do not support this usage before that.
This means that they can initially be added to a large existing table
without checking its initial contents, but new tuples must comply to
them; a separate pass invoked by ALTER TABLE / VALIDATE can verify
existing data and ensure it complies with the constraint, at which point
it is marked validated and becomes a normal part of the table ecosystem.
An non-validated CHECK constraint is ignored in the planner for
constraint_exclusion purposes; when validated, cached plans are
recomputed so that partitioning starts working right away.
This patch also enables domains to have unvalidated CHECK constraints
attached to them as well by way of ALTER DOMAIN / ADD CONSTRAINT / NOT
VALID, which can later be validated with ALTER DOMAIN / VALIDATE
CONSTRAINT.
Thanks to Thom Brown, Dean Rasheed and Jaime Casanova for the various
reviews, and Robert Hass for documentation wording improvement
suggestions.
This patch was sponsored by Enova Financial.
The existence of a btree opclass accepting composite types caused us to
assume that every composite type is sortable. This isn't true of course;
we need to check if the column types are all sortable. There was logic
for this for the case of array comparison (ie, check that the element
type is sortable), but we missed the point for rowtypes. Per Teodor's
report of an ANALYZE failure for an unsortable composite type.
Rather than just add some more ad-hoc logic for this, I moved knowledge of
the issue into typcache.c. The typcache will now only report out array_eq,
record_cmp, and friends as usable operators if the array or composite type
will work with those functions.
Unfortunately we don't have enough info to do this for anonymous RECORD
types; in that case, just assume it will work, and take the runtime failure
as before if it doesn't.
This patch might be a candidate for back-patching at some point, but
given the lack of complaints from the field, I'd rather just test it in
HEAD for now.
Note: most of the places touched in this patch will need further work
when we get around to supporting hashing of record types.
This patch is almost entirely cosmetic --- mostly cleaning up a lot of
neglected comments, and fixing code layout problems in places where the
patch made lines too long and then pgindent did weird things with that.
I did find a bug-of-omission in equalTupleDescs().
Since collation is effectively an argument, not a property of the function,
FmgrInfo is really the wrong place for it; and this becomes critical in
cases where a cached FmgrInfo is used for varying purposes that might need
different collation settings. Fix by passing it in FunctionCallInfoData
instead. In particular this allows a clean fix for bug #5970 (record_cmp
not working). This requires touching a bit more code than the original
method, but nobody ever thought that collations would not be an invasive
patch...
The previous functions of assign hooks are now split between check hooks
and assign hooks, where the former can fail but the latter shouldn't.
Aside from being conceptually clearer, this approach exposes the
"canonicalized" form of the variable value to guc.c without having to do
an actual assignment. And that lets us fix the problem recently noted by
Bernd Helmle that the auto-tune patch for wal_buffers resulted in bogus
log messages about "parameter "wal_buffers" cannot be changed without
restarting the server". There may be some speed advantage too, because
this design lets hook functions avoid re-parsing variable values when
restoring a previous state after a rollback (they can store a pre-parsed
representation of the value instead). This patch also resolves a
longstanding annoyance about custom error messages from variable assign
hooks: they should modify, not appear separately from, guc.c's own message
about "invalid parameter value".
I'm not sure these have any non-cosmetic implications, but I'm not sure
they don't, either. In particular, ensure the CaseTestExpr generated
by transformAssignmentIndirection to represent the base target column
carries the correct collation, because parse_collate.c won't fix that.
Tweak lsyscache.c API so that we can get the appropriate collation
without an extra syscache lookup.
In nearly all cases, the caller already knows the correct collation, and
in a number of places, the value the caller has handy is more correct than
the default for the type would be. (In particular, this patch makes it
significantly less likely that eval_const_expressions will result in
changing the exposed collation of an expression.) So an internal lookup
is both expensive and wrong.
Opening a catcache's index could require reading from that cache's own
catalog, which of course would acquire AccessShareLock on the catalog.
So the original coding here risks locking index before heap, which could
deadlock against another backend trying to get exclusive locks in the
normal order. Because InitCatCachePhase2 is only called when a backend
has to start up without a relcache init file, the deadlock was seldom seen
in the field. (And by the same token, there's no need to worry about any
performance disadvantage; so not much point in trying to distinguish
exactly which catalogs have the risk.)
Bug report, diagnosis, and patch by Nikhil Sontakke. Additional commentary
by me. Back-patch to all supported branches.
This patch implements data-modifying WITH queries according to the
semantics that the updates all happen with the same command counter value,
and in an unspecified order. Therefore one WITH clause can't see the
effects of another, nor can the outer query see the effects other than
through the RETURNING values. And attempts to do conflicting updates will
have unpredictable results. We'll need to document all that.
This commit just fixes the code; documentation updates are waiting on
author.
Marko Tiikkaja and Hitoshi Harada
This adds collation support for columns and domains, a COLLATE clause
to override it per expression, and B-tree index support.
Peter Eisentraut
reviewed by Pavel Stehule, Itagaki Takahiro, Robert Haas, Noah Misch
Foreign tables are a core component of SQL/MED. This commit does
not provide a working SQL/MED infrastructure, because foreign tables
cannot yet be queried. Support for foreign table scans will need to
be added in a future patch. However, this patch creates the necessary
system catalog structure, syntax support, and support for ancillary
operations such as COMMENT and SECURITY LABEL.
Shigeru Hanada, heavily revised by Robert Haas
The contents of an unlogged table are WAL-logged; thus, they are not
available on standby servers and are truncated whenever the database
system enters recovery. Indexes on unlogged tables are also unlogged.
Unlogged GiST indexes are not currently supported.
This commit replaces pg_class.relistemp with pg_class.relpersistence;
and also modifies the RangeVar node type to carry relpersistence rather
than istemp. It also removes removes rd_istemp from RelationData and
instead performs the correct computation based on relpersistence.
For clarity, we add three new macros: RelationNeedsWAL(),
RelationUsesLocalBuffers(), and RelationUsesTempNamespace(), so that we
can clarify the purpose of each check that previous depended on
rd_istemp.
This is intended as infrastructure for the upcoming unlogged tables
patch, as well as for future possible work on global temporary tables.
This is a heavily revised version of builtin_knngist_core-0.9. The
ordering operators are no longer mixed in with actual quals, which would
have confused not only humans but significant parts of the planner.
Instead, ordering operators are carried separately throughout planning and
execution.
Since the API for ambeginscan and amrescan functions had to be changed
anyway, this commit takes the opportunity to rationalize that a bit.
RelationGetIndexScan no longer forces a premature index_rescan call;
instead, callers of index_beginscan must call index_rescan too. Aside from
making the AM-side initialization logic a bit less peculiar, this has the
advantage that we do not make a useless extra am_rescan call when there are
runtime key values. AMs formerly could not assume that the key values
passed to amrescan were actually valid; now they can.
Teodor Sigaev and Tom Lane
Formerly we looked up the operators associated with each index (caching
them in relcache) and then the planner looked up the btree opfamily
containing such operators in order to build the btree-centric pathkey
representation that describes the index's sort order. This is quite
pointless for btree indexes: we might as well just use the index's opfamily
information directly. That saves syscache lookup cycles during planning,
and furthermore allows us to eliminate the relcache's caching of operators
altogether, which may help in reducing backend startup time.
I added code to plancat.c to perform the same type of double lookup
on-the-fly if it's ever faced with a non-btree amcanorder index AM.
If such a thing actually becomes interesting for production, we should
replace that logic with some more-direct method for identifying the
corresponding btree opfamily; but it's not worth spending effort on now.
There is considerably more to do pursuant to my recent proposal to get rid
of sort-operator-based representations of sort orderings, but this patch
grabs some of the low-hanging fruit. I'll look at the remainder of that
work after the current commitfest.
This commit adds columns amoppurpose and amopsortfamily to pg_amop, and
column amcanorderbyop to pg_am. For the moment all the entries in
amcanorderbyop are "false", since the underlying support isn't there yet.
Also, extend the CREATE OPERATOR CLASS/ALTER OPERATOR FAMILY commands with
[ FOR SEARCH | FOR ORDER BY sort_operator_family ] clauses to allow the new
columns of pg_amop to be populated, and create pg_dump support for dumping
that information.
I also added some documentation, although it's perhaps a bit premature
given that the feature doesn't do anything useful yet.
Teodor Sigaev, Robert Haas, Tom Lane
In the previous coding, we simply issued ALTER SEQUENCE RESTART commands,
which do not roll back on error. This meant that an error between
truncating and committing left the sequences out of sync with the table
contents, with potentially bad consequences as were noted in a Warning on
the TRUNCATE man page.
To fix, create a new storage file (relfilenode) for a sequence that is to
be reset due to RESTART IDENTITY. If the transaction aborts, we'll
automatically revert to the old storage file. This acts just like a
rewriting ALTER TABLE operation. A penalty is that we have to take
exclusive lock on the sequence, but since we've already got exclusive lock
on its owning table, that seems unlikely to be much of a problem.
The interaction of this with usual nontransactional behaviors of sequence
operations is a bit weird, but it's hard to see what would be completely
consistent. Our choice is to discard cached-but-unissued sequence values
both when the RESTART is executed, and at rollback if any; but to not touch
the currval() state either time.
In passing, move the sequence reset operations to happen before not after
any AFTER TRUNCATE triggers are fired. The previous ordering was not
logically sensible, but was forced by the need to minimize inconsistency
if the triggers caused an error. Transactional rollback is a much better
solution to that.
Patch by Steve Singer, rather heavily adjusted by me.
The core of this patch is hash_array() and associated typcache
infrastructure, which works just about exactly like the existing support
for array comparison.
In addition I did some work to ensure that the planner won't think that an
array type is hashable unless its element type is hashable, and similarly
for sorting. This includes adding a datatype parameter to op_hashjoinable
and op_mergejoinable, and adding an explicit "hashable" flag to
SortGroupClause. The lack of a cross-check on the element type was a
pre-existing bug in mergejoin support --- but it didn't matter so much
before, because if you couldn't sort the element type there wasn't any good
alternative to failing anyhow. Now that we have the alternative of hashing
the array type, there are cases where we can avoid a failure by being picky
at the planner stage, so it's time to be picky.
The issue of exactly how to combine the per-element hash values to produce
an array hash is still open for discussion, but the rest of this is pretty
solid, so I'll commit it as-is.
After much expenditure of effort, we've got this to the point where the
performance penalty is pretty minimal in typical cases.
Andrew Dunstan, reviewed by Brendan Jurd, Dean Rasheed, and Tom Lane
This patch eliminates various bizarre behaviors caused by sloppy thinking
about the difference between a domain type and its underlying array type.
In particular, the operation of updating one element of such an array
has to be considered as yielding a value of the underlying array type,
*not* a value of the domain, because there's no assurance that the
domain's CHECK constraints are still satisfied. If we're intending to
store the result back into a domain column, we have to re-cast to the
domain type so that constraints are re-checked.
For similar reasons, such a domain can't be blindly matched to an ANYARRAY
polymorphic parameter, because the polymorphic function is likely to apply
array-ish operations that could invalidate the domain constraints. For the
moment, we just forbid such matching. We might later wish to insert an
automatic downcast to the underlying array type, but such a change should
also change matching of domains to ANYELEMENT for consistency.
To ensure that all such logic is rechecked, this patch removes the original
hack of setting a domain's pg_type.typelem field to match its base type;
the typelem will always be zero instead. In those places where it's really
okay to look through the domain type with no other logic changes, use the
newly added get_base_element_type function in place of get_element_type.
catversion bumped due to change in pg_type contents.
Per bug #5717 from Richard Huxton and subsequent discussion.
SI invalidation events, rather than indirectly through the relcache.
In the previous coding, we had to flush a composite-type typcache entry
whenever we discarded the corresponding relcache entry. This caused problems
at least when testing with RELCACHE_FORCE_RELEASE, as shown in recent report
from Jeff Davis, and might result in real-world problems given the kind of
unexpected relcache flush that that test mechanism is intended to model.
The new coding decouples relcache and typcache management, which is a good
thing anyway from a structural perspective. The cost is that we have to
search the typcache linearly to find entries that need to be flushed. There
are a couple of ways we could avoid that, but at the moment it's not clear
it's worth any extra trouble, because the typcache contains very few entries
in typical operation.
Back-patch to 8.2, the same as some other recent fixes in this general area.
The patch could be carried back to 8.0 with some additional work, but given
that it's only hypothetical whether we're fixing any problem observable in
the field, it doesn't seem worth the work now.
This allows us to reliably remove all leftover temporary relation
files on cluster startup without reference to system catalogs or WAL;
therefore, we no longer include temporary relations in XLOG_XACT_COMMIT
and XLOG_XACT_ABORT WAL records.
Since these changes require including a backend ID in each
SharedInvalSmgrMsg, the size of the SharedInvalidationMessage.id
field has been reduced from two bytes to one, and the maximum number
of connections has been reduced from INT_MAX / 4 to 2^23-1. It would
be possible to remove these restrictions by increasing the size of
SharedInvalidationMessage by 4 bytes, but right now that doesn't seem
like a good trade-off.
Review by Jaime Casanova and Tom Lane.
path that specifies useTemp, but there is no active temp schema in the
current session. (This can happen if the path was saved during a transaction
that created a temp schema and was later rolled back.) For existing callers
it's sufficient to ignore the useTemp flag in this case, though we might
later want to offer an option to create a fresh temp schema. So far as I can
tell this is just an Assert failure: in a non-assert build, the code would
push a zero onto the new search path, which is useless but not very harmful.
Per bug report from Heikki.
Back-patch to 8.3; prior versions don't have this code.
Tom Lane