One would have needed out-of-tree code to observe the defects. Remove
unreferenced fields instead of completing their support functions.
Since in-tree code can't reach _readIntoClause(), no catversion bump.
This makes the header more consistent with the surroundings, with
declarations associated to a given file grouped together.
Author: Daniel Gustafsson
Discussion: https://postgr.es/m/20190608012439.GB7228@paquier.xyz
There were a number of issues in the recent commits which include typos,
code and comments mismatch, leftover function declarations. Fix them.
Reported-by: Alexander Lakhin
Author: Alexander Lakhin, Amit Kapila and Amit Langote
Reviewed-by: Amit Kapila
Discussion: https://postgr.es/m/ef0c0232-0c1d-3a35-63d4-0ebd06e31387@gmail.com
The file has been introduced in src/backend/libpq/ as of b0b39f72, but
all backend-side headers of libpq are located in src/include/libpq/.
Note that the identification path on top of the file referred to
src/include/libpq/ from the start.
Author: Michael Paquier
Reviewed-by: Stephen Frost
Discussion: https://postgr.es/m/20190607043415.GE1736@paquier.xyz
We used the same slot to store a tuple from the index, and to store a
tuple from the table. That's not OK. It worked with the heap, because
heapam_getnextslot() stores a HeapTuple to the slot, and doesn't care how
large the tts_values/nulls arrays are. But when I played with a toy table
AM implementation that used a virtual tuple, it caused memory overruns.
In the passing, tidy up comments on the ioss_PscanLen fields.
The defined callback definitions have been using references to heap for
a couple of variables and comments. This makes the whole interface more
consistent by using "table" which is more generic.
A variable storing index information was misspelled as well.
Author: Michael Paquier
Discussion: https://postgr.es/m/20190601190946.GB1905@paquier.xyz
Interestingly only C++ compilers have, so far, complained about this
odd forward declaration. This originated when IndexBuildCallback was
defined in another file, but now is completely unnecessary (but was
wrong before too, cpluspluscheck just wouldn't have noticed).
Reported-By: Tom Lane
Discussion: https://postgr.es/m/53941.1559239260@sss.pgh.pa.us
Some of the wrapper functions didn't match the callback names. Many of
them due to staying "consistent" with historic naming of the wrapped
functionality. We decided that for most cases it's more important to
be for tableam to be consistent going forward, than with the past.
The one exception is beginscan/endscan/... because it'd have looked
odd to have systable_beginscan/endscan/... with a different naming
scheme, and changing the systable_* APIs would have caused way too
much churn (including breaking a lot of external users).
Author: Ashwin Agrawal, with some small additions by Andres Freund
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CALfoeiugyrXZfX7n0ORCa4L-m834dzmaE8eFdbNR6PMpetU4Ww@mail.gmail.com
This is still using the 2.0 version of pg_bsd_indent.
I thought it would be good to commit this separately,
so as to document the differences between 2.0 and 2.1 behavior.
Discussion: https://postgr.es/m/16296.1558103386@sss.pgh.pa.us
The original coding of this view relied on a correlated IN sub-query.
Our planner is not very bright about correlated sub-queries, and even
if it were, there's no way for it to know that the output of
pg_get_publication_tables() is duplicate-free, making the de-duplicating
semantics of IN unnecessary. Hence, rewrite as a LATERAL sub-query.
This provides circa 100X speedup for me with a few hundred published
tables (the whole regression database), and things would degrade as
roughly O(published_relations * all_relations) beyond that.
Because the rules.out expected output changes, force a catversion bump.
Ordinarily we might not want to do that post-beta1; but we already know
we'll be doing a catversion bump before beta2 to fix pg_statistic_ext
issues, so it's pretty much free to fix it now instead of waiting for v13.
Per report and fix suggestion from PegoraroF10.
Discussion: https://postgr.es/m/1551385426763-0.post@n3.nabble.com
That leads to unsatisfied external references if the C compiler fails
to elide unused static functions. Apparently, we have no buildfarm
members building HEAD that have that issue ... but such compilers still
exist in the wild. Need to do something about that.
In passing, fix Berkeley-era typo in comment.
Discussion: https://postgr.es/m/27054.1558533367@sss.pgh.pa.us
The comment for SNAPSHOT_SELF was unfortunately explaining
SNAPSHOT_DIRTY, as reported by Sergei. Also expand a few comments, and
include a few more comments from heapam_visibility.c, so they're in an
AM independent place.
Reported-By: Sergei Kornilov
Author: Andres Freund
Discussion: https://postgr.es/m/9152241558192351@sas1-d856b3d759c7.qloud-c.yandex.net
Before this commit, when ANALYZE was run on a table and serializable
was used (either by virtue of an explicit BEGIN TRANSACTION ISOLATION
LEVEL SERIALIZABLE, or default_transaction_isolation being set to
serializable) a null pointer dereference lead to a crash.
The analyze scan doesn't need a snapshot (nor predicate locking), but
before this commit a scan only contained information about being a
bitmap or sample scan.
Refactor the option passing to the scan_begin callback to use a
bitmask instead. Alternatively we could have added a new boolean
parameter, but that seems harder to read. Even before this issue
various people (Heikki, Tom, Robert) suggested doing so.
These changes don't change the scan APIs outside of tableam. The flags
argument could be exposed, it's not necessary to fix this
problem. Also the wrapper table_beginscan* functions encapsulate most
of that complexity.
After these changes fixing the bug is trivial, just don't acquire
predicate lock for analyze style scans. That was already done for
bitmap heap scans. Add an assert that a snapshot is passed when
acquiring the predicate lock, so this kind of bug doesn't require
running with serializable.
Also add a comment about sample scans currently requiring predicate
locking the entire relation, that previously wasn't remarked upon.
Reported-By: Joe Wildish
Author: Andres Freund
Discussion:
https://postgr.es/m/4EA80A20-E9BF-49F1-9F01-5B66CAB21453@elusive.cxhttps://postgr.es/m/20190411164947.nkii4gaeilt4bui7@alap3.anarazel.dehttps://postgr.es/m/20190518203102.g7peu2fianukjuxm@alap3.anarazel.de
Instead add a tableam callback to do so. To avoid adding per
validation overhead, pass a scan to tuple_tid_valid. In heap's case
we'd otherwise incurred a RelationGetNumberOfBlocks() call for each
tid - which'd have added noticable overhead to nodeTidscan.c.
Author: Andres Freund
Reviewed-By: Ashwin Agrawal
Discussion: https://postgr.es/m/20190515185447.gno2jtqxyktylyvs@alap3.anarazel.de
Previously various parts of the code routed size requests through
RelationGetNumberOfBlocks[InFork]. That works if md.c is used by the
AM, but not otherwise.
Add a tableam callback to return the size of the table. As not every
AM will use postgres' BLCKSZ, have it return bytes, and have
RelationGetNumberOfBlocksInFork() round the byte size up into blocks.
To allow code outside of the AM to determine the actual relation size
map InvalidForkNumber the total size of a relation, as not every AM
might just need the postgres defined forks.
A few users of RelationGetNumberOfBlocks() ought to be converted away
from that. One case, the use of it to determine whether a tid is
valid, will be fixed in a follow up commit. Others will have to wait
for v13.
Author: Andres Freund
Discussion: https://postgr.es/m/20190423225201.3bbv6tbqzkb5w7cw@alap3.anarazel.de
Previously, gen_partprune_steps() always built executor pruning steps
using all suitable clauses, including those containing PARAM_EXEC
Params. This meant that the pruning steps were only completely safe
for executor run-time (scan start) pruning. To prune at executor
startup, we had to ignore the steps involving exec Params. But this
doesn't really work in general, since there may be logic changes
needed as well --- for example, pruning according to the last operator's
btree strategy is the wrong thing if we're not applying that operator.
The rules embodied in gen_partprune_steps() and its minions are
sufficiently complicated that tracking their incremental effects in
other logic seems quite impractical.
Short of a complete redesign, the only safe fix seems to be to run
gen_partprune_steps() twice, once to create executor startup pruning
steps and then again for run-time pruning steps. We can save a few
cycles however by noting during the first scan whether we rejected
any clauses because they involved exec Params --- if not, we don't
need to do the second scan.
In support of this, refactor the internal APIs in partprune.c to make
more use of passing information in the GeneratePruningStepsContext
struct, rather than as separate arguments.
This is, I hope, the last piece of our response to a bug report from
Alan Jackson. Back-patch to v11 where this code came in.
Discussion: https://postgr.es/m/FAD28A83-AC73-489E-A058-2681FA31D648@tvsquared.com
The original placement of this module in src/fe_utils/ is ill-considered,
because several src/common/ modules have dependencies on it, meaning that
libpgcommon and libpgfeutils now have mutual dependencies. That makes it
pointless to have distinct libraries at all. The intended design is that
libpgcommon is lower-level than libpgfeutils, so only dependencies from
the latter to the former are acceptable.
We already have the precedent that fe_memutils and a couple of other
modules in src/common/ are frontend-only, so it's not stretching anything
out of whack to treat logging.c as a frontend-only module in src/common/.
To the extent that such modules help provide a common frontend/backend
environment for the rest of common/ to use, it's a reasonable design.
(logging.c does not yet provide an ereport() emulation, but one can
dream.)
Hence, move these files over, and revert basically all of the build-system
changes made by commit cc8d41511. There are no places that need to grow
new dependencies on libpgcommon, further reinforcing the idea that this
is the right solution.
Discussion: https://postgr.es/m/a912ffff-f6e4-778a-c86a-cf5c47a12933@2ndquadrant.com
As we descend the GiST tree during insertion, we modify any downlinks on
the way down to include the new tuple we're about to insert (if they don't
cover it already). Modifying an existing downlink might cause an internal
page to split, if the new downlink tuple is larger than the old one. If
that happens, we need to back up to the parent and re-choose a page to
insert to. We used to detect that situation, thanks to the NSN-LSN
interlock normally used to detect concurrent page splits, but that got
broken by commit 9155580fd5. With that commit, we now use a dummy constant
LSN value for every page during index build, so the LSN-NSN interlock no
longer works. I thought that was OK because there can't be any other
backends modifying the index during index build, but missed that the
insertion itself can modify the page we're inserting to. The consequence
was that we would sometimes insert the new tuple to an incorrect page, one
whose downlink doesn't cover the new tuple.
To fix, add a flag to the stack that keeps track of the state while
descending tree, to indicate that a page was split, and that we need to
retry the descend from the parent.
Thomas Munro first reported that the contrib/intarray regression test was
failing occasionally on the buildfarm after commit 9155580fd5. The failure
was intermittent, because the gistchoose() function is not deterministic,
and would only occasionally create the right circumstances for this bug to
cause the failure.
Patch by Anastasia Lubennikova, with some changes by me to make it work
correctly also when the internal page split also causes the "grandparent"
to be split.
Discussion: https://www.postgresql.org/message-id/CA%2BhUKGJRzLo7tZExWfSbwM3XuK7aAK7FhdBV0FLkbUG%2BW0v0zg%40mail.gmail.com
The term "item pointer" should not be used to refer to ItemIdData
variables, since that is needlessly ambiguous. Only
ItemPointerData/ItemPointer variables should be called item pointers.
To fix, establish the convention that ItemIdData variables should always
be referred to either as "item identifiers" or "line pointers". The
term "item identifier" already predominates in docs and translatable
messages, and so should be the preferred alternative there.
Discussion: https://postgr.es/m/CAH2-Wz=c=MZQjUzde3o9+2PLAPuHTpVZPPdYxN=E4ndQ2--8ew@mail.gmail.com
We long ago decided to design the shared PgBackendStatus data structure to
minimize the cost of writing status updates, which means that writers just
have to increment the st_changecount field twice. That isn't hooked into
any sort of resource management mechanism, which means that if something
were to throw error between the two increments, the st_changecount field
would be left odd indefinitely. That would cause readers to lock up.
Now, since it's also a bad idea to leave the field odd for longer than
absolutely necessary (because readers will spin while we have it set),
the expectation was that we'd treat these segments like spinlock critical
sections, with only short, more or less straight-line, code in them.
That was fine as originally designed, but commit 9029f4b37 broke it
by inserting a significant amount of non-straight-line code into
pgstat_bestart(), code that is very capable of throwing errors, not to
mention taking a significant amount of time during which readers will spin.
We have a report from Neeraj Kumar of readers actually locking up, which
I suspect was due to an encoding conversion error in X509_NAME_to_cstring,
though conceivably it was just a garden-variety OOM failure.
Subsequent commits have loaded even more dubious code into pgstat_bestart's
critical section (and commit fc70a4b0d deserves some kind of booby prize
for managing to miss the critical section entirely, although the negative
consequences seem minimal given that the PgBackendStatus entry should be
seen by readers as inactive at that point).
The right way to fix this mess seems to be to compute all these values
into a local copy of the process' PgBackendStatus struct, and then just
copy the data back within the critical section proper. This plan can't
be implemented completely cleanly because of the struct's heavy reliance
on out-of-line strings, which we must initialize separately within the
critical section. But still, the critical section is far smaller and
safer than it was before.
In hopes of forestalling future errors of the same ilk, rename the
macros for st_changecount management to make it more apparent that
the writer-side macros create a critical section. And to prevent
the worst consequences if we nonetheless manage to mess it up anyway,
adjust those macros so that they really are a critical section, ie
they now bump CritSectionCount. That doesn't add much overhead, and
it guarantees that if we do somehow throw an error while the counter
is odd, it will lead to PANIC and a database restart to reset shared
memory.
Back-patch to 9.5 where the problem was introduced.
In HEAD, also fix an oversight in commit b0b39f72b: it failed to teach
pgstat_read_current_status to copy st_gssstatus data from shared memory to
local memory. Hence, subsequent use of that data within the transaction
would potentially see changing data that it shouldn't see.
Discussion: https://postgr.es/m/CAPR3Wj5Z17=+eeyrn_ZDG3NQGYgMEOY6JV6Y-WRRhGgwc16U3Q@mail.gmail.com
The description of the lock type for speculative insertions was
incorrect, being copy-pasted from another one.
As discussed, also move the description for all the fields of lock tag
types from the structure listing lock tag types to the set of macros
setting each LOCKTAG.
Author: John Naylor
Discussion: https://postgr.es/m/CACPNZCtA0-ybaC4fFfaDq_8p_TUOLvGxZH9Dm-=TMHZJarBa7Q@mail.gmail.com
The right way for IsCatalogRelation/Class to behave is to return true
for OIDs less than FirstBootstrapObjectId (not FirstNormalObjectId),
without any of the ad-hoc fooling around with schema membership.
The previous code was wrong because (1) it claimed that
information_schema tables were not catalog relations but their toast
tables were, which is silly; and (2) if you dropped and recreated
information_schema, which is a supported operation, the behavior
changed. That's even sillier. With this definition, "catalog
relations" are exactly the ones traceable to the postgres.bki data,
which seems like what we want.
With this simplification, we don't actually need access to the pg_class
tuple to identify a catalog relation; we only need its OID. Hence,
replace IsCatalogClass with "IsCatalogRelationOid(oid)". But keep
IsCatalogRelation as a convenience function.
This allows fixing some arguably-wrong semantics in contrib/sepgsql and
ReindexRelationConcurrently, which were using an IsSystemNamespace test
where what they really should be using is IsCatalogRelationOid. The
previous coding failed to protect toast tables of system catalogs, and
also was not on board with the general principle that user-created tables
do not become catalogs just by virtue of being renamed into pg_catalog.
We can also get rid of a messy hack in ReindexMultipleTables.
While we're at it, also rename IsSystemNamespace to IsCatalogNamespace,
because the previous name invited confusion with the more expansive
semantics used by IsSystemRelation/Class.
Also improve the comments in catalog.c.
There are a few remaining places in replication-related code that are
special-casing OIDs below FirstNormalObjectId. I'm inclined to think
those are wrong too, and if there should be any special case it should
just extend to FirstBootstrapObjectId. But first we need to debate
whether a FOR ALL TABLES publication should include information_schema.
Discussion: https://postgr.es/m/21697.1557092753@sss.pgh.pa.us
Discussion: https://postgr.es/m/15150.1557257111@sss.pgh.pa.us
This commit adds new parameter to VACUUM command, TRUNCATE,
which specifies that VACUUM should attempt to truncate off
any empty pages at the end of the table and allow the disk space
for the truncated pages to be returned to the operating system.
This parameter, if specified, overrides the vacuum_truncate
reloption. If neither the reloption nor the VACUUM option is
used, the default is true, as before.
Author: Fujii Masao
Reviewed-by: Julien Rouhaud, Masahiko Sawada
Discussion: https://postgr.es/m/CAD21AoD+qtrSDL=GSma4Wd3kLYLeRC0hPna-YAdkDeV4z156vg@mail.gmail.com
This feature was using a process local map to track the first few blocks
in the relation. The map was reset each time we get the block with enough
freespace. It was discussed that it would be better to track this map on
a per-relation basis in relcache and then invalidate the same whenever
vacuum frees up some space in the page or when FSM is created. The new
design would be better both in terms of API design and performance.
List of commits reverted, in reverse chronological order:
06c8a5090e Improve code comments in b0eaa4c51b.
13e8643bfc During pg_upgrade, conditionally skip transfer of FSMs.
6f918159a9 Add more tests for FSM.
9c32e4c350 Clear the local map when not used.
29d108cdec Update the documentation for FSM behavior..
08ecdfe7e5 Make FSM test portable.
b0eaa4c51b Avoid creation of the free space map for small heap relations.
Discussion: https://postgr.es/m/20190416180452.3pm6uegx54iitbt5@alap3.anarazel.de
In the wake of commit f912d7dec, RelationSetIndexList isn't used any
more. It was always a horrid wart, so getting rid of it is very nice.
We can also convert rd_indexvalid back to a plain boolean.
Discussion: https://postgr.es/m/28926.1556664156@sss.pgh.pa.us
The message type for temp files and for checksum failures were missing
from the union. Due to the coding style used there was no compiler error
when this happend. So change the code to actively use the union thereby
producing a compiler error if the same mistake happens again, suggested
by Tom Lane.
Author: Julien Rouhaud
Reported-By: Tomas Vondra
Discussion: https://postgr.es/m/20190430163328.zd4rrlnbvgaqlcdz@development
Most of these stem from d25f519107 "tableam: relation creation, VACUUM
FULL/CLUSTER, SET TABLESPACE.".
1) To pass data to the relation_set_new_filenode()
RelationSetNewRelfilenode() was made to update RelationData.rd_rel
directly. That's not OK however, as it makes the relcache entries
temporarily inconsistent. Which among other scenarios is a problem
if a REINDEX targets an index on pg_class - the
CatalogTupleUpdate() in RelationSetNewRelfilenode(). Presumably
that was introduced because other places in the code do so - while
those aren't "good practice" they don't appear to be actively
buggy (e.g. because system tables may not be targeted).
I (Andres) should have caught this while reviewing and signficantly
evolving the code in that commit, mea culpa.
Fix that by instead passing in the new RelFileNode as separate
argument to relation_set_new_filenode() and rely on the relcache to
update the catalog entry. Also revert that the
RelationMapUpdateMap() call was changed to immediate, and undo some
other more unnecessary changes.
2) Document that the relation_set_new_filenode cannot rely on the
whole relcache entry to be valid. It might be worthwhile to
refactor the code to never have to rely on that, but given the way
heap_create() is currently coded, that'd be a large change.
3) ATExecSetTableSpace() shouldn't do FlushRelationBuffers() itself. A
table AM might not use shared buffers at all. Move to
index_copy_data() and heapam_relation_copy_data().
4) heapam_relation_set_new_filenode() previously sometimes accessed
rel->rd_rel->relpersistence rather than the `persistence`
argument. Code movement mistake.
5) Previously heapam_relation_set_new_filenode() re-opened the smgr
relation to create the init for, if necesary. Instead have
RelationCreateStorage() return the SMgrRelation and use it to
create the init fork.
6) Add a note about the danger of modifying the relcache directly to
ATExecSetTableSpace() - it's currently not a bug because there's a
check ERRORing for catalog tables.
Regression tests and assertion improvements that together trigger the
bug described in 1) will be added in a later commit, as there is a
related bug on all branches.
Reported-By: Michael Paquier
Diagnosed-By: Tom Lane and Andres Freund
Author: Andres Freund
Reviewed-By: Tom Lane
Discussion: https://postgr.es/m/20190418011430.GA19133@paquier.xyz
This is quite unsafe, even for the case of ereport(FATAL) where we won't
return control to the interrupted code, and despite this code's use of
a flag to restrict the areas where we'd try to do it. It's possible
for example that we interrupt malloc or free while that's holding a lock
that's meant to protect against cross-thread interference. Then, any
attempt to do malloc or free within ereport() will result in a deadlock,
preventing the walreceiver process from exiting in response to SIGTERM.
We hypothesize that this explains some hard-to-reproduce failures seen
in the buildfarm.
Hence, get rid of the immediate-exit code in WalRcvShutdownHandler,
as well as the logic associated with WalRcvImmediateInterruptOK.
Instead, we need to take care that potentially-blocking operations
in the walreceiver's data transmission logic (libpqwalreceiver.c)
will respond reasonably promptly to the process's latch becoming
set and then call ProcessWalRcvInterrupts. Much of the needed code
for that was already present in libpqwalreceiver.c. I refactored
things a bit so that all the uses of PQgetResult use latch-aware
waiting, but didn't need to do much more.
These changes should be enough to ensure that libpqwalreceiver.c
will respond promptly to SIGTERM whenever it's waiting to receive
data. In principle, it could block for a long time while waiting
to send data too, and this patch does nothing to guard against that.
I think that that hazard is mostly theoretical though: such blocking
should occur only if we fill the kernel's data transmission buffers,
and we don't generally send enough data to make that happen without
waiting for input. If we find out that the hazard isn't just
theoretical, we could fix it by using PQsetnonblocking, but that
would require more ticklish changes than I care to make now.
This is a bug fix, but it seems like too big a change to push into
the back branches without much more testing than there's time for
right now. Perhaps we'll back-patch once we have more confidence
in the change.
Patch by me; thanks to Thomas Munro for review.
Discussion: https://postgr.es/m/20190416070119.GK2673@paquier.xyz
Run renumber_oids.pl to move high-numbered OIDs down, as per pre-beta
tasks specified by RELEASE_CHANGES. (The only change is 8394 -> 3428.)
Also run reformat_dat_file.pl while I'm here.
While looking at the reformat diffs, I chanced to notice that type
jsonpath had typsend and typreceive = '-', which surely is not the
intention given that jsonpath_send and jsonpath_recv exist.
Fix that. It's safe to assume that these functions have never been
tested :-(. I didn't try, but somebody should.
Commit ca4103025d left a few loose ends. The most important one
(broken pg_dump output) is already fixed by virtue of commit
3b23552ad8, but some things remained:
* When ALTER TABLE rewrites tables, the indexes must remain in the
tablespace they were originally in. This didn't work because
index recreation during ALTER TABLE runs manufactured SQL (yuck),
which runs afoul of default_tablespace in competition with the parent
relation tablespace. To fix, reset default_tablespace to the empty
string temporarily, and add the TABLESPACE clause as appropriate.
* Setting a partitioned rel's tablespace to the database default is
confusing; if it worked, it would direct the partitions to that
tablespace regardless of default_tablespace. But in reality it does
not work, and making it work is a larger project. Therefore, throw
an error when this condition is detected, to alert the unwary.
Add some docs and tests, too.
Author: Álvaro Herrera
Discussion: https://postgr.es/m/CAKJS1f_1c260nOt_vBJ067AZ3JXptXVRohDVMLEBmudX1YEx-A@mail.gmail.com
This allows table AMs that don't need these horizons. This was already
documented in the tableam relation_set_new_filenode callback, but an
assert prevented if from actually working (the test AM code contained
the change itself). Defang the asserts in the general code, and move
the stronger ones into heap AM.
Relatedly, after CLUSTER/VACUUM, we'd always assign a relfrozenxid /
relminmxid. Change the table_relation_copy_for_cluster() interface to
allow the AM to overwrite the horizons that get set on the pg_class
entry. This'd also in the future allow AMs like heap to compute a
relfrozenxid during rewrite that's the table's actual minimum rather
than a pre-determined value. Arguably it'd have been better to move
the whole computation / setting of those values into the callback, but
it seems likely that for other reasons it'd be better to be able to
use one value to vacuum/cluster multiple tables (e.g. a toast's
horizon shouldn't be different than the table's).
Reported-By: Heikki Linnakangas
Author: Andres Freund
Discussion: https://postgr.es/m/9a7fb9cc-2419-5db7-8840-ddc10c93f122@iki.fi
Commit dd299df8 made nbtree treat heap TID as a tiebreaker column,
establishing the principle that there is only one correct location (page
and page offset number) for every index tuple, no matter what.
Insertions of tuples into non-unique indexes proceed as if heap TID
(scan key's scantid) is just another user-attribute value, but
insertions into unique indexes are more delicate. The TID value in
scantid must initially be omitted to ensure that the unique index
insertion visits every leaf page that duplicates could be on. The
scantid is set once again after unique checking finishes successfully,
which can force _bt_findinsertloc() to step right one or more times, to
locate the leaf page that the new tuple must be inserted on.
Stepping right within _bt_findinsertloc() was assumed to occur no more
frequently than stepping right within _bt_check_unique(), but there was
one important case where that assumption was incorrect: inserting a
"duplicate" with NULL values. Since _bt_check_unique() didn't do any
real work in this case, it wasn't appropriate for _bt_findinsertloc() to
behave as if it was finishing off a conventional unique insertion, where
any existing physical duplicate must be dead or recently dead.
_bt_findinsertloc() might have to grovel through a substantial portion
of all of the leaf pages in the index to insert a single tuple, even
when there were no dead tuples.
To fix, treat insertions of tuples with NULLs into a unique index as if
they were insertions into a non-unique index: never unset scantid before
calling _bt_search() to descend the tree, and bypass _bt_check_unique()
entirely. _bt_check_unique() is no longer responsible for incoming
tuples with NULL values.
Discussion: https://postgr.es/m/CAH2-Wzm08nr+JPx4jMOa9CGqxWYDQ-_D4wtPBiKghXAUiUy-nQ@mail.gmail.com
Up to now, DefineIndex() was responsible for adding attnotnull constraints
to the columns of a primary key, in any case where it hadn't been
convenient for transformIndexConstraint() to mark those columns as
is_not_null. It (or rather its minion index_check_primary_key) did this
by executing an ALTER TABLE SET NOT NULL command for the target table.
The trouble with this solution is that if we're creating the index due
to ALTER TABLE ADD PRIMARY KEY, and the outer ALTER TABLE has additional
sub-commands, the inner ALTER TABLE's operations executed at the wrong
time with respect to the outer ALTER TABLE's operations. In particular,
the inner ALTER would perform a validation scan at a point where the
table's storage might be inconsistent with its catalog entries. (This is
on the hairy edge of being a security problem, but AFAICS it isn't one
because the inner scan would only be interested in the tuples' null
bitmaps.) This can result in unexpected failures, such as the one seen
in bug #15580 from Allison Kaptur.
To fix, let's remove the attempt to do SET NOT NULL from DefineIndex(),
reducing index_check_primary_key's role to verifying that the columns are
already not null. (It shouldn't ever see such a case, but it seems wise
to keep the check for safety.) Instead, make transformIndexConstraint()
generate ALTER TABLE SET NOT NULL subcommands to be executed ahead of
the ADD PRIMARY KEY operation in every case where it can't force the
column to be created already-not-null. This requires only minor surgery
in parse_utilcmd.c, and it makes for a much more satisfying spec for
transformIndexConstraint(): it's no longer having to take it on faith
that someone else will handle addition of NOT NULL constraints.
To make that work, we have to move the execution of AT_SetNotNull into
an ALTER pass that executes ahead of AT_PASS_ADD_INDEX. I moved it to
AT_PASS_COL_ATTRS, and put that after AT_PASS_ADD_COL to avoid failure
when the column is being added in the same command. This incidentally
fixes a bug in the only previous usage of AT_PASS_COL_ATTRS, for
AT_SetIdentity: it didn't work either for a newly-added column.
Playing around with this exposed a separate bug in ALTER TABLE ONLY ...
ADD PRIMARY KEY for partitioned tables. The intent of the ONLY modifier
in that context is to prevent doing anything that would require holding
lock for a long time --- but the implied SET NOT NULL would recurse to
the child partitions, and do an expensive validation scan for any child
where the column(s) were not already NOT NULL. To fix that, invent a
new ALTER subcommand AT_CheckNotNull that just insists that a child
column be already NOT NULL, and apply that, not AT_SetNotNull, when
recursing to children in this scenario. This results in a slightly laxer
definition of ALTER TABLE ONLY ... SET NOT NULL for partitioned tables,
too: that command will now work as long as all children are already NOT
NULL, whereas before it just threw up its hands if there were any
partitions.
In passing, clean up the API of generateClonedIndexStmt(): remove a
useless argument, ensure that the output argument is not left undefined,
update the header comment.
A small side effect of this change is that no-such-column errors in ALTER
TABLE ADD PRIMARY KEY now produce a different message that includes the
table name, because they are now detected by the SET NOT NULL step which
has historically worded its error that way. That seems fine to me, so
I didn't make any effort to avoid the wording change.
The basic bug #15580 is of very long standing, and these other bugs
aren't new in v12 either. However, this is a pretty significant change
in the way ALTER TABLE ADD PRIMARY KEY works. On balance it seems best
not to back-patch, at least not till we get some more confidence that
this patch has no new bugs.
Patch by me, but thanks to Jie Zhang for a preliminary version.
Discussion: https://postgr.es/m/15580-d1a6de5a3d65da51@postgresql.org
Discussion: https://postgr.es/m/1396E95157071C4EBBA51892C5368521017F2E6E63@G08CNEXMBPEKD02.g08.fujitsu.local
Robert Haas
Noah Misch
Michael Paquier
Amit Kapila
Heikki Linnakangas
Peter Eisentraut
David Rowley
Andres Freund
Tom Lane
Peter Geoghegan
Etsuro Fujita
Fujii Masao
Magnus Hagander
Alvaro Herrera