The logic in ATExecDropColumn that rejects dropping partition key
columns is quite an inadequate defense, because it doesn't execute
in cases where a column needs to be dropped due to cascade from
something that only the column, not the whole partitioned table,
depends on. That leaves us with a badly broken partitioned table;
even an attempt to load its relcache entry will fail.
We really need to have explicit pg_depend entries that show that the
column can't be dropped without dropping the whole table. Hence,
add those entries. In v12 and HEAD, bump catversion to ensure that
partitioned tables will have such entries. We can't do that in
released branches of course, so in v10 and v11 this patch affords
protection only to partitioned tables created after the patch is
installed. Given the lack of field complaints (this bug was found
by fuzz-testing not by end users), that's probably good enough.
In passing, fix ATExecDropColumn and ATPrepAlterColumnType
messages to be more specific about which partition key column
they're complaining about.
Per report from Manuel Rigger. Back-patch to v10 where partitioned
tables were added.
Discussion: https://postgr.es/m/CA+u7OA4JKCPFrdrAbOs7XBiCyD61XJxeNav4LefkSmBLQ-Vobg@mail.gmail.com
Discussion: https://postgr.es/m/31920.1562526703@sss.pgh.pa.us
Change the defaults for the pg_hba.conf generated by initdb to "peer"
for local (if supported, else "md5") and "md5" for host.
(Changing from "md5" to SCRAM is left as a separate exercise.)
"peer" is currently not supported on AIX, HP-UX, and Windows. Users
on those operating systems will now either have to provide a password
to initdb or choose a different authentication method when running
initdb.
Reviewed-by: Julien Rouhaud <rjuju123@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/bec17f0a-ddb1-8b95-5e69-368d9d0a3390%40postgresql.org
When we already know the length that we're going to append, then it
makes sense to use appendBinaryStringInfo instead of
appendStringInfoString so that the append can be performed with a simple
memcpy() using a known length rather than having to first perform a
strlen() call to obtain the length.
Discussion: https://postgr.es/m/CAKJS1f8+FRAM1s5+mAa3isajeEoAaicJ=4e0WzrH3tAusbbiMQ@mail.gmail.com
Some code could get confused when certain catalog state involving both
identity and serial sequences was present, perhaps during an attempt
to upgrade the latter to the former. Specifically, dropping the
default of a serial column maintains the ownership of the sequence by
the column, and so it would then be possible to afterwards make the
column an identity column that would now own two sequences. This
causes the code that looks up the identity sequence to error out,
making the new identity column inoperable until the ownership of the
previous sequence is released.
To fix this, make the identity sequence lookup only consider sequences
with the appropriate dependency type for an identity sequence, so it
only ever finds one (unless something else is broken). In the above
example, the old serial sequence would then be ignored. Reorganize
the various owned-sequence-lookup functions a bit to make this
clearer.
Reported-by: Laurenz Albe <laurenz.albe@cybertec.at>
Discussion: https://www.postgresql.org/message-id/flat/470c54fc8590be4de0f41b0d295fd6390d5e8a6c.camel@cybertec.at
In nodeAppend.c and nodeMergeAppend.c there were some foreach loops which
looped over the list of subplans and only performed any work if the
subplan index was found in a Bitmapset. With the old linked list
implementation of List, this form made sense as accessing the Nth list
element was O(N). However, thanks to 1cff1b95a we now have array-based
lists, so accessing the Nth element has become O(1).
Here we make the most of the O(1) lookups and just loop over the set
members of the Bitmapset with bms_next_member(). This performs slightly
better when a small number of the list items are in the Bitmapset. Micro
benchmarks show that when the Bitmapset contains all or most of the list
items then the new code is ever so slightly slower. In practice, the cost
is so small that it's drowned out by various other things such as locking
the relations belonging to each subplan, etc.
The primary goal here is to leave better code examples around which benefit
better from the new list implementation.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CAKJS1f8ZcsLVgkF4wOfRyMYTcPgLFiUAOedFC+U2vK_aFZk-BA@mail.gmail.com
3373c7155 changed how we determine EquivalenceClasses for relations and
added an Assert to ensure all relations mentioned in each EC's ec_relids
was a RELOPT_BASEREL. However, the join removal code may remove a LEFT
JOIN and since it does not clean up EC members belonging to the removed
relations it can leave RELOPT_DEADREL rels in ec_relids.
Fix this by adjusting the Assert to allow RELOPT_DEADREL rels too.
Reported-by: sqlsmith via Andreas Seltenreich
Discussion: https://postgr.es/m/87y30r8sls.fsf@ansel.ydns.eu
Coverity complained about this code, apparently because it uses a local
array of size FUNC_MAX_ARGS without a guard that the input argument list
is no longer than that. (Not sure why it complained today, since this
code's been the same for a long time; possibly it re-analyzed everything
the List API change touched?)
Rather than add a guard, though, let's just get rid of the local array
altogether. It was only there to avoid list_nth() calls, and those are
no longer expensive.
Several buildfarm members have been complaining about that with gcc,
like jacana. Weirdly enough, Visual Studio's compilers do not find this
issue.
Author: Michael Paquier
Reviewed-by: Andrew Dunstan
Discussion: https://postgr.es/m/20190719050830.GK1859@paquier.xyz
Previously in order to determine which ECs a relation had members in, we
had to loop over all ECs stored in PlannerInfo's eq_classes and check if
ec_relids mentioned the relation. For the most part, this was fine, as
generally, unless queries were fairly complex, the overhead of performing
the lookup would have not been that significant. However, when queries
contained large numbers of joins and ECs, the overhead to find the set of
classes matching a given set of relations could become a significant
portion of the overall planning effort.
Here we allow a much more efficient method to access the ECs which match a
given relation or set of relations. A new Bitmapset field in RelOptInfo
now exists to store the indexes into PlannerInfo's eq_classes list which
each relation is mentioned in. This allows very fast lookups to find all
ECs belonging to a single relation. When we need to lookup ECs belonging
to a given pair of relations, we can simply bitwise-AND the Bitmapsets from
each relation and use the result to perform the lookup.
We also take the opportunity to write a new implementation of
generate_join_implied_equalities which makes use of the new indexes.
generate_join_implied_equalities_for_ecs must remain as is as it can be
given a custom list of ECs, which we can't easily determine the indexes of.
This was originally intended to fix the performance penalty of looking up
foreign keys matching a join condition which was introduced by 100340e2d.
However, we're speeding up much more than just that here.
Author: David Rowley, Tom Lane
Reviewed-by: Tom Lane, Tomas Vondra
Discussion: https://postgr.es/m/6970.1545327857@sss.pgh.pa.us
Solely relying on a relation's reltuples/relpages estimate to size the
Bloom filters used by amcheck verification makes verification less
effective when the estimates are very stale. In extreme cases,
verification options that use Bloom filters internally could be totally
ineffective, without users receiving any clear indication that certain
types of corruption might easily be missed.
To fix, use RelationGetNumberOfBlocks() instead of relpages to size the
downlink block Bloom filter. Use the same RelationGetNumberOfBlocks()
value to derive a minimum size for the heapallindexed Bloom filter,
rather than completely trusting reltuples. Verification will still be
reasonably effective when the projected/estimated number of Bloom filter
elements is at least 1/5 of the final number of elements, which is
assured by the new sizing logic.
Reported-By: Alexander Korotkov
Discussion: https://postgr.es/m/CAH2-Wzk0ke2J42KrNYBKu0Xovjy-sU5ub7PWjgpbsKdAQcL4OA@mail.gmail.com
Backpatch: 11-, where downlink/heapallindexed verification were added.
The current extended statistics code was a bit confused which collation
to use. When building the statistics, the collations defined as default
for the data types were used (since commit 5e0928005). The MCV code was
however using the column collations for MCV serialization, and then
DEFAULT_COLLATION_OID when computing estimates. So overall the code was
using all three possible options, inconsistently.
This uses the column colation everywhere - this makes it consistent with
what 5e0928005 did for regular stats. We however do not track the
collations in a catalog, because we can derive them from column-level
information. This may need to change in the future, e.g. after allowing
statistics on expressions.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
Backpatch-to: 12
The examine_opclause_expression function needs to return information on
which side of the operator we found the Var, but the variable was called
"isgt" which is rather misleading (it assumes the operator is either
less-than or greater-than, but it may be equality or something else).
Other places in the planner use a variable called "varonleft" for this
purpose, so just adopt the same convention here.
The code also assumed we don't care about this flag for equality, as
(Var = Const) and (Const = Var) should be the same thing. But that does
not work for cross-type operators, in which case we need to pass the
parameters to the procedure in the right order. So just use the same
code for all types of expressions.
This means we don't need to care about the selectivity estimation
function anymore, at least not in this code. We should only get the
supported cases here (thanks to statext_is_compatible_clause).
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
Backpatch-to: 12
Absorb commit e5e04c962a5d12eebbf867ca25905b3ccc34cbe0 from upstream
IANA code, in hopes of silencing warnings from MSVC about negating
a bool value.
Discussion: https://postgr.es/m/20190719035347.GJ1859@paquier.xyz
16828d5 has improved ALTER TABLE so as a column addition does not
require a rewrite for a non-NULL default with constant expressions, but
one spot in the documentation did not get updated consistently.
The documentation also now clarifies the fact that this does not apply
if the expression is volatile, where a table rewrite is still required.
Reported-by: Daniel Westermann
Author: Ian Barwick
Reviewed-by: Michael Paquier, Daniel Westermann
Discussion: https://postgr.es/m/DB6PR0902MB2184C7D5645CF15D75EB7957D2CF0@DB6PR0902MB2184.eurprd09.prod.outlook.com
Backpatch-through: 11
The existing facility of vacuumdb to handle parallel connections into a
given database with an authentication set is moved to a common file in
src/bin/scripts/, named scripts_parallel.c. This introduces a set of
routines to initialize, wait and terminate a set of connections,
simplifying a bit the code of vacuumdb on the way. More routines
related to result handling and database connection are moved to
common.c.
The initial plan is to use that for reindexdb, but it could be applied
to other tools like clusterdb.
While on it, clean up a set of variables "progname" which were defined
as routine arguments for error messages. Since most of the callers have
switched to pg_log_error() and such there is no need for this variable.
Author: Julien Rouhaud
Reviewed-by: Michael Paquier, Álvaro Herrera
Discussion: https://postgr.es/m/CAOBaU_YrnH_Jqo46NhaJ7uRBiWWEcS40VNRQxgFbqYo9kApUsg@mail.gmail.com
The values 'infinity' and '-infinity' are a part of the DATE type
itself, so a bound of the date 'infinity' is not the same as an
unbounded/infinite range. However, it is still wrong to try to
canonicalize such values, because adding or subtracting one has no
effect. Fix by treating 'infinity' and '-infinity' the same as
unbounded ranges for the purposes of canonicalization (but not other
purposes).
Backpatch to all versions because it is inconsistent with the
documented behavior. Note that this could be an incompatibility for
applications relying on the behavior contrary to the documentation.
Author: Laurenz Albe
Reviewed-by: Thomas Munro
Discussion: https://postgr.es/m/77f24ea19ab802bc9bc60ddbb8977ee2d646aec1.camel%40cybertec.at
Backpatch-through: 9.4
Commit 857f9c36cd, which taught nbtree VACUUM to avoid unnecessary
index scans, bumped the nbtree version number from 2 to 3, while adding
the ability for nbtree indexes to be upgraded on-the-fly. Various
assertions that assumed that an nbtree index was always on version 2 had
to be changed to accept any supported version (version 2 or 3 on
Postgres 11).
However, a few assertions were missed in the initial commit, all of
which were in code paths that cache a local copy of the metapage
metadata, where the index had been expected to be on the current version
(no longer version 2) as a generic sanity check. Rather than simply
update the assertions, follow-up commit 0a64b45152 intentionally made
the metapage caching code update the per-backend cached metadata version
without changing the on-disk version at the same time. This could even
happen when the planner needed to determine the height of a B-Tree for
costing purposes. The assertions only fail on Postgres v12 when
upgrading from v10, because they were adjusted to use the authoritative
shared memory metapage by v12's commit dd299df8.
To fix, remove the cache-only upgrade mechanism entirely, and update the
assertions themselves to accept any supported version (go back to using
the cached version in v12). The fix is almost a full revert of commit
0a64b45152 on the v11 branch.
VACUUM only considers the authoritative metapage, and never bothers with
a locally cached version, whereas everywhere else isn't interested in
the metapage fields that were added by commit 857f9c36cd. It seems
unlikely that this bug has affected any user on v11.
Reported-By: Christoph Berg
Bug: #15896
Discussion: https://postgr.es/m/15896-5b25e260fdb0b081%40postgresql.org
Backpatch: 11-, where VACUUM was taught to avoid unnecessary index scans.
Now that commit fec0778c8 drew a clear line between public and private
fields in SPITupleTable, it seems pretty silly that the count of valid
tuples isn't on the public side of that line. The reason why not was
that there wasn't such a count. For reasons lost in the mists of time,
spi.c preferred to keep a count of remaining free entries in the array.
But that seems pretty pointless: it's unlike the way we handle similar
code everywhere else, and it involves extra subtractions that surely
outweigh having to do a comparison rather than test-for-zero to check
for array-full.
Hence, rearrange so that this code does the expansible array logic
the same as everywhere else, with a count of valid entries alongside
the allocated array length. And document the count as public.
I looked for core-code callers where it would make sense to start
relying on tuptable->numvals rather than the separate SPI_processed
variable. Right now there don't seem to be places where it'd be
a win to do so without more code restructuring than I care to
undertake today. In principle, though, having SPITupleTables be
fully self-contained should be helpful down the line.
Discussion: https://postgr.es/m/16852.1563395722@sss.pgh.pa.us
When evaluating clauses on a multivariate MCV list, we build a bitmap
tracking how the clauses match each item of the MCV list. When updating
the bitmap we need to consider the current value (tracking how the item
matches preceding clauses), match for the current clause and whether the
clauses are connected by AND or OR.
Until now the logic was copied on every place updating the bitmap, which
was not quite readable. So just move it to a separate function and call
it where needed.
Backpatch to 12, where the code was introduced. While not a bugfix, this
should make maintenance and future backpatches easier.
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
There were two issues in how the extended statistics handled NULL values
in opclauses. Firstly, the code was oblivious to the possibility that
Const may be NULL (constisnull=true) in which case the constvalue is
undefined. We need to treat this as a mismatch, and not call the proc.
Secondly, the MCV item itself may contain NULL values too - the code
already did check that, and updated the match bitmap accordingly, but
failed to ensure we won't call the operator procedure anyway. It did
work for AND-clauses, because in that case false in the bitmap stops
evaluation of further clauses. But for OR-clauses ir was not easy to
get incorrect estimates or even trigger a crash.
This fixes both issues by extending the existing check so that it looks
at constisnull too, and making sure it skips calling the procedure.
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
We expect opclauses to have exactly one Var and one Const, but the code
was checking the Const by calling is_pseudo_constant_clause() which is
incorrect - we need a proper constant.
Fixed by using plain IsA(x,Const) to check type of the node. We need to
do these checks in two places, so move it into a separate function that
can be called in both places.
Reported by Andreas Seltenreich, based on crash reported by sqlsmith.
Backpatch to v12, where this code was introduced.
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
Backpatch-to: 12
The TYPECACHE_GT_OPR is not needed (it used to be in older version of
the MCV code), but the compiler failed to detect this as the result was
used in a fmgr_info() call, populating a FmgrInfo entry.
Backpatch to v12, where this code was introduced.
Discussion: https://postgr.es/m/8736jdhbhc.fsf%40ansel.ydns.eu
Backpatch-to: 12
The documentation mentioned that data checksums cannot be changed after
initialization, which is not true as pg_checksums can do that with its
--enable option introduced in v12. This simply removes the sentence
telling so.
Reported-by: Basil Bourque
Author: Michael Paquier
Reviewed-by: Daniel Gustafsson
Discussion: https://postgr.es/m/15909-e9d74271f1647472@postgresql.org
Backpatch-through: 12
A large fraction of this diff is just due to upstream's somewhat
random decision to rename a bunch of internal variables and struct
fields. However, there is an interesting new feature in zic:
it's grown a "-b slim" option that emits zone files without 32-bit
data and other backwards-compatibility hacks. We should consider
whether we wish to enable that.
The fields that we consider public are "tupdesc" and "vals", which
historically are in the middle of the struct. Move them to the front
(this should be perfectly safe to do in HEAD) and add comments to make
it quite clear which fields are public or not.
Also adjust spi.sgml's documentation of the struct to match.
That doc had bit-rotted somewhat, as it was missing some fields.
(Arguably we should just remove all the private fields from the docs,
but for now I refrained.)
Daniel Gustafsson, reviewed by Fabien Coelho
Discussion: https://postgr.es/m/0D19F836-B743-4340-B6A2-F148CA3DD1F0@yesql.se
At least on Fedora and RHEL, it's not in the same RPM that's needed
for building sepgsql itself. Today is the second or third time I've
had to rediscover how to install that, so let's document it this time.
The aggregate-order difference explained in my previous commit
turns out to also affect the order of log entries emitted in the
contrib/sepgsql regression test. Per buildfarm.
Discussion: https://postgr.es/m/21272.1563318411@sss.pgh.pa.us
Formerly, lcons was about the same speed as lappend, but with the new
List implementation, that's not so; with a long List, data movement
imposes an O(N) cost on lcons and list_delete_first, but not lappend.
Hence, invent list_delete_last with semantics parallel to
list_delete_first (but O(1) cost), and change various places to use
lappend and list_delete_last where this can be done without much
violence to the code logic.
There are quite a few places that construct result lists using lcons not
lappend. Some have semantic rationales for that; I added comments about
it to a couple that didn't have them already. In many such places though,
I think the coding is that way only because back in the dark ages lcons
was faster than lappend. Hence, switch to lappend where this can be done
without causing semantic changes.
In ExecInitExprRec(), this results in aggregates and window functions that
are in the same plan node being executed in a different order than before.
Generally, the executions of such functions ought to be independent of
each other, so this shouldn't result in visibly different query results.
But if you push it, as one regression test case does, you can show that
the order is different. The new order seems saner; it's closer to
the order of the functions in the query text. And we never documented
or promised anything about this, anyway.
Also, in gistfinishsplit(), don't bother building a reverse-order list;
it's easy now to iterate backwards through the original list.
It'd be possible to go further towards removing uses of lcons and
list_delete_first, but it'd require more extensive logic changes,
and I'm not convinced it's worth it. Most of the remaining uses
deal with queues that probably never get long enough to be worth
sweating over. (Actually, I doubt that any of the changes in this
patch will have measurable performance effects either. But better
to have good examples than bad ones in the code base.)
Patch by me, thanks to David Rowley and Daniel Gustafsson for review.
Discussion: https://postgr.es/m/21272.1563318411@sss.pgh.pa.us
Potential future SMGR implementations may not want to create
tablespace directories when creating an SMGR relation. Move that
logic to mdcreate(). Move the initialization of md-specific
data structures from smgropen() to a new callback mdopen().
Author: Thomas Munro
Reviewed-by: Shawn Debnath (as part of an earlier patch set)
Discussion: https://postgr.es/m/CA%2BhUKG%2BOZqOiOuDm5tC5DyQZtJ3FH4%2BFSVMqtdC4P1atpJ%2Bqhg%40mail.gmail.com
This should lappend the OIDs, not lcons them; the existing code produced
a list in reversed order. This is harmless for single-key FKs or FKs
where all the key columns are of the same type, which probably explains
how it went unnoticed. But if those conditions are not met,
ATAddForeignKeyConstraint would make the wrong decision about whether an
existing FK needs to be revalidated. I think it would almost always err
in the safe direction by revalidating a constraint that didn't need it.
You could imagine scenarios where the pfeqop check was fooled by
swapping the types of two FK columns in one ALTER TABLE, but that case
would probably be rejected by other tests, so it might be impossible to
get to the worst-case scenario where an FK should be revalidated and
isn't. (And even then, it's likely to be fine, unless there are weird
inconsistencies in the equality behavior of the replacement types.)
However, this is a performance bug at least.
Noted while poking around to see whether lcons calls could be converted
to lappend.
This bug is old, dating to commit cb3a7c2b9, so back-patch to all
supported branches.
It seems worth getting rid of these functions because they require the
caller to retain a ListCell pointer into a List that it's modifying,
which is a dangerous practice with the new List implementation.
(The only other List-modifying function that takes a ListCell pointer
as input is list_delete_cell, which nowadays is preferentially used
via the constrained API foreach_delete_current.)
There was only one remaining caller of these functions after commit
2f5b8eb5a, and that was some fairly ugly GEQO code that can be much
more clearly expressed using a list-index variable and list_insert_nth.
Hence, rewrite that code, and remove the functions.
Discussion: https://postgr.es/m/26193.1563228600@sss.pgh.pa.us
heap.c and relcache.c contained nearly identical copies of logic
to insert OIDs into an OID list while preserving the list's OID
ordering (and rejecting duplicates, in one case but not the other).
The comments argue that this is faster than qsort for small numbers
of OIDs, which is at best unproven, and seems even less likely to be
true now that lappend_cell_oid has to move data around. In any case
it's ugly and hard-to-follow code, and if we do have a lot of OIDs
to consider, it's O(N^2).
Hence, replace with simply lappend'ing OIDs to a List, then list_sort
the completed List, then remove adjacent duplicates if necessary.
This is demonstrably O(N log N) and it's much simpler for the
callers. It's possible that this would be somewhat inefficient
if there were a very large number of duplicates, but that seems
unlikely in the existing usage.
This adds list_deduplicate_oid and list_oid_cmp infrastructure
to list.c. I didn't bother with equivalent functionality for
integer or pointer Lists, but such could always be added later
if we find a use for it.
Discussion: https://postgr.es/m/26193.1563228600@sss.pgh.pa.us
In the wake of commit 1cff1b95a, the obvious way to sort a List
is to apply qsort() directly to the array of ListCells. list_qsort
was building an intermediate array of pointers-to-ListCells, which
we no longer need, but getting rid of it forces an API change:
the comparator functions need to do one less level of indirection.
Since we're having to touch the callers anyway, let's do two additional
changes: sort the given list in-place rather than making a copy (as
none of the existing callers have any use for the copying behavior),
and rename list_qsort to list_sort. It was argued that the old name
exposes more about the implementation than it should, which I find
pretty questionable, but a better reason to rename it is to be sure
we get the attention of any external callers about the need to fix
their comparator functions.
While we're at it, change four existing callers of qsort() to use
list_sort instead; previously, they all had local reinventions
of list_qsort, ie build-an-array-from-a-List-and-qsort-it.
(There are some other places where changing to list_sort perhaps
would be worthwhile, but they're less obviously wins.)
Discussion: https://postgr.es/m/29361.1563220190@sss.pgh.pa.us
This is numbered take 7, and addresses a set of issues around:
- Fixes for typos and incorrect reference names.
- Removal of unneeded comments.
- Removal of unreferenced functions and structures.
- Fixes regarding variable name consistency.
Author: Alexander Lakhin
Discussion: https://postgr.es/m/10bfd4ac-3e7c-40ab-2b2e-355ed15495e8@gmail.com
The logic just added by commit e3899ffd falls back on a 50:50 page split
in the event of a new item that's just to the right of our provisional
"many duplicates" split point. Fix a comment that incorrectly claimed
that the new item had to be just to the left of our provisional split
point.
Backpatch: 12-, just like commit e3899ffd.
Specific ever-decreasing insertion patterns could cause successive
unbalanced nbtree page splits. Problem cases involve a large group of
duplicates to the left, and ever-decreasing insertions to the right.
To fix, detect the situation by considering the newitem offset before
performing a split using nbtsplitloc.c's "many duplicates" strategy. If
the new item was inserted just to the right of our provisional "many
duplicates" split point, infer ever-decreasing insertions and fall back
on a 50:50 (space delta optimal) split. This seems to barely affect
cases that already had acceptable space utilization.
An alternative fix also seems possible. Instead of changing
nbtsplitloc.c split choice logic, we could instead teach _bt_truncate()
to generate a new value for new high keys by interpolating from the
lastleft and firstright key values. That would certainly be a more
elegant fix, but it isn't suitable for backpatching.
Discussion: https://postgr.es/m/CAH2-WznCNvhZpxa__GqAa1fgQ9uYdVc=_apArkW2nc-K3O7_NA@mail.gmail.com
Backpatch: 12-, where the nbtree page split enhancements were introduced.
Originally, Postgres Lists were a more or less exact reimplementation of
Lisp lists, which consist of chains of separately-allocated cons cells,
each having a value and a next-cell link. We'd hacked that once before
(commit d0b4399d8) to add a separate List header, but the data was still
in cons cells. That makes some operations -- notably list_nth() -- O(N),
and it's bulky because of the next-cell pointers and per-cell palloc
overhead, and it's very cache-unfriendly if the cons cells end up
scattered around rather than being adjacent.
In this rewrite, we still have List headers, but the data is in a
resizable array of values, with no next-cell links. Now we need at
most two palloc's per List, and often only one, since we can allocate
some values in the same palloc call as the List header. (Of course,
extending an existing List may require repalloc's to enlarge the array.
But this involves just O(log N) allocations not O(N).)
Of course this is not without downsides. The key difficulty is that
addition or deletion of a list entry may now cause other entries to
move, which it did not before.
For example, that breaks foreach() and sister macros, which historically
used a pointer to the current cons-cell as loop state. We can repair
those macros transparently by making their actual loop state be an
integer list index; the exposed "ListCell *" pointer is no longer state
carried across loop iterations, but is just a derived value. (In
practice, modern compilers can optimize things back to having just one
loop state value, at least for simple cases with inline loop bodies.)
In principle, this is a semantics change for cases where the loop body
inserts or deletes list entries ahead of the current loop index; but
I found no such cases in the Postgres code.
The change is not at all transparent for code that doesn't use foreach()
but chases lists "by hand" using lnext(). The largest share of such
code in the backend is in loops that were maintaining "prev" and "next"
variables in addition to the current-cell pointer, in order to delete
list cells efficiently using list_delete_cell(). However, we no longer
need a previous-cell pointer to delete a list cell efficiently. Keeping
a next-cell pointer doesn't work, as explained above, but we can improve
matters by changing such code to use a regular foreach() loop and then
using the new macro foreach_delete_current() to delete the current cell.
(This macro knows how to update the associated foreach loop's state so
that no cells will be missed in the traversal.)
There remains a nontrivial risk of code assuming that a ListCell *
pointer will remain good over an operation that could now move the list
contents. To help catch such errors, list.c can be compiled with a new
define symbol DEBUG_LIST_MEMORY_USAGE that forcibly moves list contents
whenever that could possibly happen. This makes list operations
significantly more expensive so it's not normally turned on (though it
is on by default if USE_VALGRIND is on).
There are two notable API differences from the previous code:
* lnext() now requires the List's header pointer in addition to the
current cell's address.
* list_delete_cell() no longer requires a previous-cell argument.
These changes are somewhat unfortunate, but on the other hand code using
either function needs inspection to see if it is assuming anything
it shouldn't, so it's not all bad.
Programmers should be aware of these significant performance changes:
* list_nth() and related functions are now O(1); so there's no
major access-speed difference between a list and an array.
* Inserting or deleting a list element now takes time proportional to
the distance to the end of the list, due to moving the array elements.
(However, it typically *doesn't* require palloc or pfree, so except in
long lists it's probably still faster than before.) Notably, lcons()
used to be about the same cost as lappend(), but that's no longer true
if the list is long. Code that uses lcons() and list_delete_first()
to maintain a stack might usefully be rewritten to push and pop at the
end of the list rather than the beginning.
* There are now list_insert_nth...() and list_delete_nth...() functions
that add or remove a list cell identified by index. These have the
data-movement penalty explained above, but there's no search penalty.
* list_concat() and variants now copy the second list's data into
storage belonging to the first list, so there is no longer any
sharing of cells between the input lists. The second argument is
now declared "const List *" to reflect that it isn't changed.
This patch just does the minimum needed to get the new implementation
in place and fix bugs exposed by the regression tests. As suggested
by the foregoing, there's a fair amount of followup work remaining to
do.
Also, the ENABLE_LIST_COMPAT macros are finally removed in this
commit. Code using those should have been gone a dozen years ago.
Patch by me; thanks to David Rowley, Jesper Pedersen, and others
for review.
Discussion: https://postgr.es/m/11587.1550975080@sss.pgh.pa.us
In order to be able to work with FullTransactionId values during replay
without increasing the size of the WAL, infer the epoch. In general we
can't do that safely, but during replay we can because we know that
nextFullXid can't advance concurrently.
Prevent frontend code from seeing this new function, due to the above
restriction. Perhaps in future it will be possible to extract the value
entirely from independent WAL records, and then this restriction can be
lifted.
Author: Thomas Munro, based on earlier code from Andres Freund
Discussion: https://postgr.es/m/CA%2BhUKG%2BmLmuDjMi6o1dxkKvGRL56Y2Rz%2BiXAcrZV03G9ZuFQ8Q%40mail.gmail.com
Michael Paquier
Tom Lane
Peter Eisentraut
David Rowley
Peter Geoghegan
Tomas Vondra
Jeff Davis
Andres Freund
Thomas Munro
Bruce Momjian