Previously we just set the seed based on process ID and start timestamp.
Both those values are directly available within the session, and can
be found out or guessed by other users too, making the session's series
of random(3) values fairly predictable. Up to now, our backend-internal
uses of random(3) haven't seemed security-critical, but commit 88bdbd3f7
added one that potentially is: when using log_statement_sample_rate, a
user might be able to predict which of his SQL statements will get logged.
To improve this situation, upgrade the per-process seed initialization
method to use pg_strong_random() if available, greatly reducing the
predictability of the initial seed value. This adds a few tens of
microseconds to process start time, but since backend startup time is
at least a couple of milliseconds, that seems an acceptable price.
This means that pg_strong_random() needs to be able to run without
reliance on any backend infrastructure, since it will be invoked
before any of that is up. It was safe for that already, but adjust
comments and #include commands to make it clearer.
Discussion: https://postgr.es/m/3859.1545849900@sss.pgh.pa.us
Previously, the SQL random() function depended on libc's random(3),
and setseed() invoked srandom(3). This results in interference between
these functions and backend-internal uses of random(3). We'd never paid
too much mind to that, but in the wake of commit 88bdbd3f7 which added
log_statement_sample_rate, the interference arguably has a security
consequence: if log_statement_sample_rate is active then an unprivileged
user could probably control which if any of his SQL commands get logged,
by issuing setseed() at the right times. That seems bad.
To fix this reliably, we need random() and setseed() to use their own
private random state variable. Standard random(3) isn't amenable to such
usage, so let's switch to pg_erand48(). It's hard to say whether that's
more or less "random" than any particular platform's version of random(3),
but it does have a wider seed value and a longer period than are required
by POSIX, so we can hope that this isn't a big downgrade. Also, we should
now have uniform behavior of random() across platforms, which is worth
something.
While at it, upgrade the per-process seed initialization method to use
pg_strong_random() if available, greatly reducing the predictability
of the initial seed value. (I'll separately do something similar for
the internal uses of random().)
In addition to forestalling the possible security problem, this has a
benefit in the other direction, which is that we can now document
setseed() as guaranteeing a reproducible sequence of random() values.
Previously, because of the possibility of internal calls of random(3),
we could not promise any such thing.
Discussion: https://postgr.es/m/3859.1545849900@sss.pgh.pa.us
Get rid of the multiplier and addend variables in favor of hard-wired
constants. Do the multiply-and-add using uint64 arithmetic, rather
than manually combining several narrower multiplications and additions.
Make _dorand48 return the full-width new random value, and have its
callers use that directly (after suitable masking) rather than
reconstructing what they need from the unsigned short[] representation.
On my machine, this is good for a nearly factor-of-2 speedup of
pg_erand48(), probably mostly from needing just one call of ldexp()
rather than three. The wins for the other functions are smaller
but measurable. While none of the existing call sites are really
performance-critical, a cycle saved is a cycle earned; and besides
the machine code is smaller this way (at least on x86_64).
Patch by me, but the original idea to optimize this by switching
to int64 arithmetic is from Fabien Coelho.
Discussion: https://postgr.es/m/1551.1546018192@sss.pgh.pa.us
POSIX specifies that jrand48() returns a signed 32-bit value (in the
range [-2^31, 2^31)), but our code was returning an unsigned 32-bit
value (in the range [0, 2^32)). This doesn't actually matter to any
existing call site, because they all cast the "long" result to int32
or uint32; but it will doubtless bite somebody in the future.
To fix, cast the arithmetic result to int32 explicitly before the
compiler widens it to long (if widening is needed).
While at it, upgrade this file's far-short-of-project-style comments.
Had there been some peer pressure to document pg_jrand48() properly,
maybe this thinko wouldn't have gotten committed to begin with.
Backpatch to v10 where pg_jrand48() was added, just in case somebody
back-patches a fix that uses it and depends on the standard behavior.
Discussion: https://postgr.es/m/17235.1545951602@sss.pgh.pa.us
The documentation of ON DELETE and ON UPDATE uses the term "action",
which is also used on the ALTER TABLE page for other purposes. This
commit renames the term to "referential_action", which is more
consistent with the SQL specification. The new term is now used on the
documentation of both CREATE TABLE and ALTER TABLE for consistency.
Reported-by: Brigitte Blanc-Lafay
Author: Lætitia Avrot
Reviewed-by: Álvaro Herrera
Discussion: https://postgr.es/m/CAB_COdiHEVVs0uB+uYCjjYUwQ4YFFekppq+Xqv6qAM8+cd42gA@mail.gmail.com
Isolation test suite of GIN predicate locking was criticized for being too slow,
especially under Valgrind. This commit is intended to accelerate it. Tests are
simplified in the following ways.
1) Amount of data is reduced. We're now close to the minimal amount of data,
which produces at least one posting tree and at least two pages of entry
tree.
2) Three isolation tests are merged into one.
3) Only one tuple is queried from posting tree. So, locking of index is the
same, but tuple locks are not propagated to relation lock. Also, it is
faster.
4) Test cases itself are simplified. Now each test case run just one INSERT
and one SELECT involving GIN, which either conflict or not.
Discussion: https://postgr.es/m/20181204000740.ok2q53nvkftwu43a%40alap3.anarazel.de
Reported-by: Andres Freund
Tested-by: Andrew Dunstan
Author: Alexander Korotkov
Backpatch-through: 11
According to README we acquire predicate locks on entry tree leafs and posting
tree roots. However, when ginFindLeafPage() is going to lock leaf in exclusive
mode, then it checks root for conflicts regardless whether it's a entry or
posting tree. Assuming that we never place predicate lock on entry tree root
(excluding corner case when root is leaf), this check is redundant. This
commit removes this check. Now, root conflict checking is controlled by
separate argument of ginFindLeafPage().
Discussion: https://postgr.es/m/CAPpHfdv7rrDyy%3DMgsaK-L9kk0AH7az0B-mdC3w3p0FSb9uoyEg%40mail.gmail.com
Author: Alexander Korotkov
Backpatch-through: 11
Inheritance trees can include temporary tables if the parent is
permanent, which makes possible the presence of multiple temporary
children from different sessions. Trying to issue a TRUNCATE on the
parent in this scenario causes a failure, so similarly to any other
queries just ignore such cases, which makes TRUNCATE work
transparently.
This makes truncation behave similarly to any other DML query working on
the parent table with queries which need to be work on the children. A
set of isolation tests is added to cover basic cases.
Reported-by: Zhou Digoal
Author: Amit Langote, Michael Paquier
Discussion: https://postgr.es/m/15565-ce67a48d0244436a@postgresql.org
Backpatch-through: 9.4
While it seems OK to not be concerned about fsync() failure for a
pre-existing signal file, it's not OK to not even check for open()
failure. This at least causes complaints from static analyzers,
and I think on some platforms passing -1 to fsync() or close() might
trigger assertion-type failures. Also add (void) casts to make clear
that we're ignoring fsync's result intentionally.
Oversights in commit 2dedf4d9a, noted by Coverity.
I made a frontend fprintf() format use %m, forgetting that that's only
safe in HEAD not the back branches; prior to 96bf88d52 and d6c55de1f,
it would work on glibc platforms but not elsewhere. Revert to using
%s ... strerror(errno) as the code did before.
We could have left HEAD as-is, but for code consistency across branches,
I chose to apply this patch there too.
Per Coverity and a few buildfarm members.
This fixes two issues with the completion of ALTER TABLE and ALTER INDEX
after SET STATISTICS is typed, trying to suggest schema objects while
the grammar only allows integers.
The tab completion of ALTER INDEX is made smarter by handling properly
more patterns. COLUMN is an optional keyword, but as no column numbers
can be suggested yet as possible input simply adjust the completion so
as no incorrect queries are generated.
Author: Michael Paquier
Reviewed-by: Tatsuro Yamada
Discussion: https://postgr.es/m/20181219092255.GC680@paquier.xyz
At the end of recovery for the post-promotion process, a new history
file is created followed by the last partial segment of the previous
timeline. Based on the timing, the archiver would first try to archive
the last partial segment and then the history file. This can delay the
detection of a new timeline taken, particularly depending on the time it
takes to transfer the last partial segment as it delays the moment the
history file of the new timeline gets archived. This can cause promoted
standbys to use the same timeline as one already taken depending on the
circumstances if multiple instances look at archives at the same
location.
This commit changes the order of archiving so as history files are
archived in priority over other file types, which reduces the likelihood
of the same timeline being taken (still not reducing the window to
zero), and it makes the archiver behave more consistently with the
startup process doing its post-promotion business.
Author: David Steele
Reviewed-by: Michael Paquier, Kyotaro Horiguchi
Discussion: https://postgr.es/m/929068cf-69e1-bba2-9dc0-e05986aed471@pgmasters.net
Backpatch-through: 9.5
COPY can skip writing WAL when loading data on a table which has been
created in the same transaction as the one loading the data, however
this cannot work on views or foreign table as this would result in
trying to flush relation files which do not exist. So disable the
optimization so as commands are able to work the same way with any
configuration of wal_level.
Tests are added to cover the different cases, which need to have
wal_level set to minimal to allow the problem to show up, and that is
not the default configuration.
Reported-by: Luis M. Carril, Etsuro Fujita
Author: Amit Langote, Michael Paquier
Reviewed-by: Etsuro Fujita
Discussion: https://postgr.es/m/15552-c64aa14c5c22f63c@postgresql.org
Backpatch-through: 10, where support for COPY on views has been added,
while v11 has added support for COPY on foreign tables.
In passing, move the list of parameters where it can be used for both
CREATE TABLE and ALTER TABLE, and reorder it alphabetically.
Author: Dagfinn Ilmari Mannsåker
Discussion: https://postgr.es/m/d8j1s77kdbb.fsf@dalvik.ping.uio.no
013ebc0a7b implements so-called GiST microvacuum. That is gistgettuple() marks
index tuples as dead when kill_prior_tuple is set. Later, when new tuple
insertion claims page space, those dead index tuples are physically deleted
from page. When this deletion is replayed on standby, it might conflict with
read-only queries. But 013ebc0a7b doesn't handle this. That may lead to
disappearance of some tuples from read-only snapshots on standby.
This commit implements resolving of conflicts between replay of GiST microvacuum
and standby queries. On the master we implement new WAL record type
XLOG_GIST_DELETE, which comprises necessary information. On stable releases
we've to be tricky to keep WAL compatibility. Information required for conflict
processing is just appended to data of XLOG_GIST_PAGE_UPDATE record. So,
PostgreSQL version, which doesn't know about conflict processing, will just
ignore that.
Reported-by: Andres Freund
Diagnosed-by: Andres Freund
Discussion: https://postgr.es/m/20181212224524.scafnlyjindmrbe6%40alap3.anarazel.de
Author: Alexander Korotkov
Backpatch-through: 9.6
The SQL spec says that sql_identifier is a domain over varchar,
but it also says that that domain is supposed to represent the set
of valid identifiers for the implementation, in particular applying
a length limit matching the implementation's identifier length limit.
We were declaring sql_identifier as just "character varying", thus
duplicating what the spec says about base type, but entirely failing
at the rest of it.
Instead, let's declare sql_identifier as a domain over type "name".
(We can drop the COLLATE "C" added by commit 6b0faf723, since that's
now implicit in "name".) With the recent improvements to name's
comparison support, there's not a lot of functional difference between
name and varchar. So although in principle this is a spec deviation,
it's a pretty minor one. And correctly enforcing PG's name length limit
is a good thing; on balance this seems closer to the intent of the spec
than what we had.
But that's all just language-lawyering. The *real* reason to do this is
that it makes sql_identifier columns exposed by information_schema views
be just direct representations of the underlying "name" catalog columns,
eliminating a semantic mismatch that was disastrous for performance of
typical queries on the information_schema. In combination with the
recent change to allow dropping no-op CoerceToDomain nodes, this allows
(for example) queries such as
select ... from information_schema.tables where table_name = 'foo';
to produce an indexscan rather than a seqscan on pg_class.
Discussion: https://postgr.es/m/CAFj8pRBUCX4LZ2rA2BbEkdD6NN59mgx+BLo1gO08Wod4RLtcTg@mail.gmail.com
information_schema output columns that are declared as being type
sql_identifier are supposed to conform to the implementation's rules
for valid identifiers, in particular the identifier length limit.
Several places potentially violated this limit by concatenating a
function's name and OID. (The OID is added to ensure name uniqueness
within a schema, since the spec doesn't expect function name overloading.)
Simply truncating the concatenation result to fit in "name" won't do,
since losing part of the OID might wind up giving non-unique results.
Instead, let's truncate the function name as necessary.
The most practical way to do that is to do it in a C function; the
information_schema.sql script doesn't have easy access to the value
of NAMEDATALEN, nor does it have an easy way to truncate on the basis
of resulting byte-length rather than number of characters.
(There are still a couple of places that cast concatenation results to
sql_identifier, but as far as I can see they are guaranteed not to produce
over-length strings, at least with the normal value of NAMEDATALEN.)
Discussion: https://postgr.es/m/23817.1545283477@sss.pgh.pa.us
For probably bogus reasons, we acquire only AccessShareLock on the
partition when we try to detach it from its parent partitioned table.
This can cause ugly things to happen if another transaction is doing
any sort of DDL to the partition concurrently.
Upgrade that lock to ShareUpdateExclusiveLock, which per discussion
seems to be the minimum needed.
Reported by Robert Haas.
Discussion: https://postgr.es/m/CA+TgmoYruJQ+2qnFLtF1xQtr71pdwgfxy3Ziy-TxV28M6pEmyA@mail.gmail.com
"$user" in a search_path string is replaced by CURRENT_USER not
SESSION_USER. (It actually was SESSION_USER in the initial implementation,
but we changed it shortly later, and evidently forgot to fix the docs to
match.)
Noted by antonov@stdpr.ru
Discussion: https://postgr.es/m/159151fb45d490c8d31ea9707e9ba99d@stdpr.ru
Using the full width of the CPU's native word size shouldn't cost
anything in typical cases. When working with large bitmapsets,
this halves the number of operations needed for many common BMS
operations. On the right sort of test case, a measurable improvement
is obtained.
David Rowley
Discussion: https://postgr.es/m/CAKJS1f9EGBd2h-VkXvb=51tf+X46zMX5T8h-KYgXEV_u2zmLUw@mail.gmail.com
When a partition is detached from its parent, we acquire locks on all
attached indexes to also detach them ... but we release those locks
immediately. This is a violation of the policy of keeping locks on user
objects to the end of the transaction. Bug introduced in 8b08f7d482.
It's unclear that there are any ill effects possible, but it's clearly
wrong nonetheless. It's likely that bad behavior *is* possible, but
mostly because the relation that the index is for is only locked with
AccessShareLock, which is an older bug that shall be fixed separately.
While touching that line of code, close the index opened with
index_open() using index_close() instead of relation_close().
No difference in practice, but let's be consistent.
Unearthed by Robert Haas.
Discussion: https://postgr.es/m/CA+TgmoYruJQ+2qnFLtF1xQtr71pdwgfxy3Ziy-TxV28M6pEmyA@mail.gmail.com
The following completion patterns are added:
- CREATE TABLE <name> with '(', OF or PARTITION OF.
- CREATE TABLE <name> OF with list of composite types.
- CREATE TABLE name (...) with PARTITION OF, WITH, TABLESPACE, ON
COMMIT (depending on the presence of a temporary table).
- CREATE TABLE ON COMMIT with actions (only for temporary tables).
Author: Dagfinn Ilmari Mannsåker
Discussion: https://postgr.es/m/d8j1s77kdbb.fsf@dalvik.ping.uio.no
The flag for IF NOT EXISTS was only being passed down in the normal
recursing case. It's been this way since originally added in 9.6 in
commit 2cd40adb85 so backpatch back to 9.6.
Now that name comparison has effectively the same behavior as text
comparison, we might as well merge the name_ops opfamily into text_ops,
allowing cross-type comparisons to be processed without forcing a
datatype coercion first. We need do little more than add cross-type
operators to make the opfamily complete, and fix one or two places
in the planner that assumed text_ops was a single-datatype opfamily.
I chose to unify hash name_ops into hash text_ops as well, since the
types have compatible hashing semantics. This allows marking the
new cross-type equality operators as oprcanhash.
(Note: this doesn't remove the name_ops opclasses, so there's no
breakage of index definitions. Those opclasses are just reparented
into the text_ops opfamily.)
Discussion: https://postgr.es/m/15938.1544377821@sss.pgh.pa.us
The "name" comparison operators now all support collations, making them
functionally equivalent to "text" comparisons, except for the different
physical representation of the datatype. They do, in fact, mostly share
the varstr_cmp and varstr_sortsupport infrastructure, which has been
slightly enlarged to handle the case.
To avoid changes in the default behavior of the datatype, set name's
typcollation to C_COLLATION_OID not DEFAULT_COLLATION_OID, so that
by default comparisons to a name value will continue to use strcmp
semantics. (This would have been the case for system catalog columns
anyway, because of commit 6b0faf723, but doing this makes it true for
user-created name columns as well. In particular, this avoids
locale-dependent changes in our regression test results.)
In consequence, tweak a couple of places that made assumptions about
collatable base types always having typcollation DEFAULT_COLLATION_OID.
I have not, however, attempted to relax the restriction that user-
defined collatable types must have that. Hence, "name" doesn't
behave quite like a user-defined type; it acts more like a domain
with COLLATE "C". (Conceivably, if we ever get rid of the need for
catalog name columns to be fixed-length, "name" could actually become
such a domain over text. But that'd be a pretty massive undertaking,
and I'm not volunteering.)
Discussion: https://postgr.es/m/15938.1544377821@sss.pgh.pa.us
Avoid repetitive calls to repalloc() when the required size of the
collector array grows more than 2x in one call. Also ensure that the
array size is a power of 2 (since palloc will probably consume a power
of 2 anyway) and doesn't start out very small (which'd likely just lead
to extra repallocs).
David Rowley, tweaked a bit by me
Discussion: https://postgr.es/m/CAKJS1f8vn-iSBE8PKeVHrnhvyjRNYCxguPFFY08QLYmjWG9hPQ@mail.gmail.com
Remove a comment from the Berkeley days claiming that nbtree must
disambiguate duplicate keys within _bt_moveright(). There is no special
care taken around duplicates within _bt_moveright(), at least since
commit 9e85183bfc removed inscrutable _bt_moveright() code to handle
pages full of duplicates.
Commit 40dae7ec53, which made the handling of interrupted nbtree page
splits more robust, removed an nbtree-specific end-of-recovery cleanup
step. This meant that it was no longer possible to complete an
interrupted page split during recovery. However, a reference to
recovery as a reason for using a NULL stack while inserting into a
parent page was missed. Remove the reference.
Remove a similar obsolete reference to recovery that was introduced much
more recently, as part of the btree fastpath optimization enhancement
that made it into Postgres 11 (commit 2b272734, and follow-up commits).
Backpatch: 11-, where the fastpath optimization was introduced.
Up to now we allowed text columns in system catalogs to use collation
"default", but that isn't really safe because it might mean something
different in template0 than it means in a database cloned from template0.
In particular, this could mean that cloned pg_statistic entries for such
columns weren't entirely valid, possibly leading to bogus planner
estimates, though (probably) not any outright failures.
In the wake of commit 5e0928005, a better solution is available: if we
label such columns with "C" collation, then their pg_statistic entries
will also use that collation and hence will be valid independently of
the database collation.
This also provides a cleaner solution for indexes on such columns than
the hack added by commit 0b28ea79c: the indexes will naturally inherit
"C" collation and don't have to be forced to use text_pattern_ops.
Also, with the planned improvement of type "name" to be collation-aware,
this policy will apply cleanly to both text and name columns.
Because of the pg_statistic angle, we should also apply this policy
to the tables in information_schema. This patch does that by adjusting
information_schema's textual domain types to specify "C" collation.
That has the user-visible effect that order-sensitive comparisons to
textual information_schema view columns will now use "C" collation
by default. The SQL standard says that the collation of those view
columns is implementation-defined, so I think this is legal per spec.
At some point this might allow for translation of such comparisons
into indexable conditions on the underlying "name" columns, although
additional work will be needed before that can happen.
Discussion: https://postgr.es/m/19346.1544895309@sss.pgh.pa.us
"rescanratio" was computed as 1 + rescanned-tuples / total-inner-tuples,
which is sensible if it's to be multiplied by total-inner-tuples or a cost
value corresponding to scanning all the inner tuples. But in reality it
was (mostly) multiplied by inner_rows or a related cost, numbers that take
into account the possibility of stopping short of scanning the whole inner
relation thanks to a limited key range in the outer relation. This'd
still make sense if we could expect that stopping short would result in a
proportional decrease in the number of tuples that have to be rescanned.
It does not, however. The argument that establishes the validity of our
estimate for that number is independent of whether we scan all of the inner
relation or stop short, and experimentation also shows that stopping short
doesn't reduce the number of rescanned tuples. So the correct calculation
is 1 + rescanned-tuples / inner_rows, and we should be sure to multiply
that by inner_rows or a corresponding cost value.
Most of the time this doesn't make much difference, but if we have
both a high rescan rate (due to lots of duplicate values) and an outer
key range much smaller than the inner key range, then the error can
be significant, leading to a large underestimate of the cost associated
with rescanning.
Per report from Vijaykumar Jain. This thinko appears to go all the way
back to the introduction of the rescan estimation logic in commit
70fba7043, so back-patch to all supported branches.
Discussion: https://postgr.es/m/CAE7uO5hMb_TZYJcZmLAgO6iD68AkEK6qCe7i=vZUkCpoKns+EQ@mail.gmail.com
The new grammar pattern of ALTER INDEX SET STATISTICS able to use column
numbers on top of the existing column names introduced by commit 5b6d13e
forgot to add support for the feature in pg_dump, so defining statistics
on index columns was missing from the dumps, potentially causing silent
planning problems with a subsequent restore.
pg_dump ought to not use column names in what it generates as these are
automatically generated by the server and could conflict with real
relation attributes with matching patterns. "expr" and "exprN", N
incremented automatically after the creation of the first one, are used
as default attribute names for index expressions, and that could easily
match what is defined in other relations, causing the dumps to fail if
some of those attributes are renamed at some point. So to avoid any
problems, the new grammar with column numbers gets used.
Reported-by: Ronan Dunklau
Author: Michael Paquier
Reviewed-by: Tom Lane, Adrien Nayrat, Amul Sul
Discussion: https://postgr.es/m/CAARsnT3UQ4V=yDNW468w8RqHfYiY9mpn2r_c5UkBJ97NAApUEw@mail.gmail.com
Backpatch-through: 11, where the new syntax has been introduced.
It appears that all platforms that have sys_siglist[] also have
strsignal(), making that fallback case in pg_strsignal() dead code.
Getting rid of it allows dropping a configure test, which seems worth
more than providing textual signal descriptions on whatever platforms
might still hypothetically have use for the fallback case.
Discussion: https://postgr.es/m/25758.1544983503@sss.pgh.pa.us
When partitioned tables were introduced, we failed to realize that by
copying the tablespace handling for other relation kinds with no
physical storage we were causing the secondary effect that their
partitions would not automatically inherit the tablespace setting. This
is surprising and unhelpful, so change it to adopt the behavior
introduced in pg11 (commit 33e6c34c32) for partitioned indexes: the
parent relation remembers the tablespace specification, which is then
used for any new partitions that don't declare one.
Because this commit changes behavior of the TABLESPACE clause for
partitioned tables (it's no longer a no-op), it is not backpatched.
Author: David Rowley, Álvaro Herrera
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/CAKJS1f9SxVzqDrGD1teosFd6jBMM0UEaa14_8mRvcWE19Tu0hA@mail.gmail.com
