Document these as "nearest integer >= argument" and "nearest integer <=
argument", which will hopefully be less confusing than the old formulation.
New wording is from Matlab via Dean Rasheed.
I changed the pg_description entries as well as the SGML docs. In the
back branches, this will only affect installations initdb'd in the future,
but it should be harmless otherwise.
Discussion: <CAEZATCW3yzJo-NMSiQs5jXNFbTsCEftZS-Og8=FvFdiU+kYuSA@mail.gmail.com>
If we ANALYZE only selected columns of a table, we should not postpone
auto-analyze because of that; other columns may well still need stats
updates. As committed, the counter is left alone if a column list is
given, whether or not it includes all analyzable columns of the table.
Per complaint from Tomasz Ostrowski.
It's been like this a long time, so back-patch to all supported branches.
Report: <ef99c1bd-ff60-5f32-2733-c7b504eb960c@ato.waw.pl>
The original intent in the stats collector was that we should not write out
stats data oftener than every PGSTAT_STAT_INTERVAL msec. Backends will not
make requests at all if they see the existing data is newer than that, and
the stats collector is supposed to disregard requests having a cutoff_time
older than its most recently written data, so that close-together requests
don't result in multiple writes. But the latter part of that got broken
in commit 187492b6c2e8cafc, so that if two backends concurrently decide
the existing stats are too old, the collector would write the data twice.
(In principle the collector's logic would still merge requests as long as
the second one arrives before we've actually written data ... but since
the message collection loop would write data immediately after processing
a single inquiry message, that never happened in practice, and in any case
the window in which it might work would be much shorter than
PGSTAT_STAT_INTERVAL.)
To fix, improve pgstat_recv_inquiry so that it checks whether the cutoff
time is too old, and doesn't add a request to the queue if so. This means
that we do not need DBWriteRequest.request_time, because the decision is
taken before making a queue entry. And that means that we don't really
need the DBWriteRequest data structure at all; an OID list of database
OIDs will serve and allow removal of some rather verbose and crufty code.
In passing, improve the comments in this area, which have been rather
neglected. Also change backend_read_statsfile so that it's not silently
relying on MyDatabaseId to have some particular value in the autovacuum
launcher process. It accidentally worked as desired because MyDatabaseId
is zero in that process; but that does not seem like a dependency we want,
especially with no documentation about it.
Although this patch is mine, it turns out I'd rediscovered a known bug,
for which Tomas Vondra had already submitted a patch that's functionally
equivalent to the non-cosmetic aspects of this patch. Thanks to Tomas
for reviewing this version.
Back-patch to 9.3 where the bug was introduced.
Prior-Discussion: <1718942738eb65c8407fcd864883f4c8@fuzzy.cz>
Patch: <4625.1464202586@sss.pgh.pa.us>
If both timeout indicators are set when we arrive at ProcessInterrupts,
we've historically just reported "lock timeout". However, some buildfarm
members have been observed to fail isolationtester's timeouts test by
reporting "lock timeout" when the statement timeout was expected to fire
first. The cause seems to be that the process is allowed to sleep longer
than expected (probably due to heavy machine load) so that the lock
timeout happens before we reach the point of reporting the error, and
then this arbitrary tiebreak rule does the wrong thing. We can improve
matters by comparing the scheduled timeout times to decide which error
to report.
I had originally proposed greatly reducing the 1-second window between
the two timeouts in the test cases. On reflection that is a bad idea,
at least for the case where the lock timeout is expected to fire first,
because that would assume that it takes negligible time to get from
statement start to the beginning of the lock wait. Thus, this patch
doesn't completely remove the risk of test failures on slow machines.
Empirically, however, the case this handles is the one we are seeing
in the buildfarm. The explanation may be that the other case requires
the scheduler to take the CPU away from a busy process, whereas the
case fixed here only requires the scheduler to not give the CPU back
right away to a process that has been woken from a multi-second sleep
(and, perhaps, has been swapped out meanwhile).
Back-patch to 9.3 where the isolationtester timeouts test was added.
Discussion: <8693.1464314819@sss.pgh.pa.us>
do_text_output_multiline() would fail (typically with a null pointer
dereference crash) if its input string did not end with a newline. Such
cases do not arise in our current sources; but it certainly could happen
in future, or in extension code's usage of the function, so we should fix
it. To fix, replace "eol += len" with "eol = text + len".
While at it, make two cosmetic improvements: mark the input string const,
and rename the argument from "text" to "txt" to dodge pgindent strangeness
(since "text" is a typedef name).
Even though this problem is only latent at present, it seems like a good
idea to back-patch the fix, since it's a very simple/safe patch and it's
not out of the realm of possibility that we might in future back-patch
something that expects sane behavior from do_text_output_multiline().
Per report from Hao Lee.
Report: <CAGoxFiFPAGyPAJLcFxTB5cGhTW2yOVBDYeqDugYwV4dEd1L_Ag@mail.gmail.com>
This patch essentially reverts commit 4c6780fd17aa43ed, in favor of a much
simpler solution for the case where the new cluster would choose to create
a TOAST table but the old cluster doesn't have one: just don't create a
TOAST table.
The existing code failed in at least two different ways if the situation
arose: (1) ALTER TABLE RESET didn't grab an exclusive lock, so that the
lock sanity check in create_toast_table failed; (2) pg_upgrade did not
provide a pg_type OID for the new toast table, so that the crosscheck in
TypeCreate failed. While both these problems were introduced by later
patches, they show that the hack being used to cause TOAST table creation
is overwhelmingly fragile (and untested). I also note that before the
TypeCreate crosscheck was added, the code would have resulted in assigning
an indeterminate pg_type OID to the toast table, possibly causing a later
OID conflict in that catalog; so that it didn't really work even when
committed.
If we simply don't create a TOAST table, there will only be a problem if
the code tries to store a tuple that's wider than a page, and field
compression isn't sufficient to get it under a page. Given that the TOAST
creation threshold is intended to be about a quarter of a page, it's very
hard to believe that cross-version differences in the do-we-need-a-toast-
table heuristic could result in an observable problem. So let's just
follow the old version's conclusion about whether a TOAST table is needed.
(If we ever do change needs_toast_table() so much that this conclusion
doesn't apply, we can devise a solution at that time, and hopefully do
it in a less klugy way than 4c6780fd17aa43ed did.)
Back-patch to 9.3, like the previous patch.
Discussion: <8110.1462291671@sss.pgh.pa.us>
Given a three-or-more-way equivalence class, such as X.Y = Y.Y = Z.Z,
it was possible for the planner to omit one of the quals needed to
enforce that all members of the equivalence class are actually equal.
This only happened in the case of a parameterized join node for two
of the relations, that is a plan tree like
Nested Loop
-> Scan X
-> Nested Loop
-> Scan Y
-> Scan Z
Filter: Z.Z = X.X
The eclass machinery normally expects to apply X.X = Y.Y when those
two relations are joined, but in this shape of plan tree they aren't
joined until the top node --- and, if the lower nested loop is marked
as parameterized by X, the top node will assume that the relevant eclass
condition(s) got pushed down into the lower node. On the other hand,
the scan of Z assumes that it's only responsible for constraining Z.Z
to match any one of the other eclass members. So one or another of
the required quals sometimes fell between the cracks, depending on
whether consideration of the eclass in get_joinrel_parampathinfo()
for the lower nested loop chanced to generate X.X = Y.Y or X.X = Z.Z
as the appropriate constraint there. If it generated the latter,
it'd erroneously suppose that the Z scan would take care of matters.
To fix, force X.X = Y.Y to be generated and applied at that join node
when this case occurs.
This is *extremely* hard to hit in practice, because various planner
behaviors conspire to mask the problem; starting with the fact that the
planner doesn't really like to generate a parameterized plan of the
above shape. (It might have been impossible to hit it before we
tweaked things to allow this plan shape for star-schema cases.) Many
thanks to Alexander Kirkouski for submitting a reproducible test case.
The bug can be demonstrated in all branches back to 9.2 where parameterized
paths were introduced, so back-patch that far.
Commit 23a41573c attempted to fix the DatumGetBool macro to ignore bits
in a Datum that are to the left of the actual bool value. But it did that
by casting the Datum to bool; and on compilers that use C99 semantics for
bool, that ends up being a whole-word test, not a 1-byte test. This seems
to be the true explanation for contrib/seg failing in VS2015. To fix, use
GET_1_BYTE() explicitly. I think in the previous patch, I'd had some idea
of not having to commit to bool being exactly 1 byte wide, but regardless
of what the compiler's bool is, boolean columns and Datums are certainly
1 byte wide.
The previous fix was (eventually) back-patched into all active versions,
so do likewise with this one.
The coverage was rather lean for cases that bind() or listen() might
return. Add entries for everything that there's a direct equivalent
for in the set of Unix errnos that elog.c has heard of.
There's no longer a need for the MSVC-version-specific code stanza that
forcibly redefines errno code symbols, because since commit 73838b52 we're
unconditionally redefining them in the stanza before this one anyway.
Now it's merely confusing and ugly, so get rid of it; and improve the
comment that explains what's going on here.
Although this is just cosmetic, back-patch anyway since I'm intending
to back-patch some less-cosmetic changes in this same hunk of code.
Fix Windows builds to report something useful rather than "could not bind
IPv4 socket: No error" when bind() fails.
Back-patch of commits d1b7d4877b9a71f4 and 22989a8e34168f57.
Discussion: <4065.1452450340@sss.pgh.pa.us>
Commit 36a35c550ac114ca turned the interface between ginPlaceToPage and
its subroutines in gindatapage.c and ginentrypage.c into a royal mess:
page-update critical sections were started in one place and finished in
another place not even in the same file, and the very same subroutine
might return having started a critical section or not. Subsequent patches
band-aided over some of the problems with this design by making things
even messier.
One user-visible resulting problem is memory leaks caused by the need for
the subroutines to allocate storage that would survive until ginPlaceToPage
calls XLogInsert (as reported by Julien Rouhaud). This would not typically
be noticeable during retail index updates. It could be visible in a GIN
index build, in the form of memory consumption swelling to several times
the commanded maintenance_work_mem.
Another rather nasty problem is that in the internal-page-splitting code
path, we would clear the child page's GIN_INCOMPLETE_SPLIT flag well before
entering the critical section that it's supposed to be cleared in; a
failure in between would leave the index in a corrupt state. There were
also assorted coding-rule violations with little immediate consequence but
possible long-term hazards, such as beginning an XLogInsert sequence before
entering a critical section, or calling elog(DEBUG) inside a critical
section.
To fix, redefine the API between ginPlaceToPage() and its subroutines
by splitting the subroutines into two parts. The "beginPlaceToPage"
subroutine does what can be done outside a critical section, including
full computation of the result pages into temporary storage when we're
going to split the target page. The "execPlaceToPage" subroutine is called
within a critical section established by ginPlaceToPage(), and it handles
the actual page update in the non-split code path. The critical section,
as well as the XLOG insertion call sequence, are both now always started
and finished in ginPlaceToPage(). Also, make ginPlaceToPage() create and
work in a short-lived memory context to eliminate the leakage problem.
(Since a short-lived memory context had been getting created in the most
common code path in the subroutines, this shouldn't cause any noticeable
performance penalty; we're just moving the overhead up one call level.)
In passing, fix a bunch of comments that had gone unmaintained throughout
all this klugery.
Report: <571276DD.5050303@dalibo.com>
Per discussion, there doesn't seem to be much value in having
NUM_SPINLOCK_SEMAPHORES set to 1024: under any scenario where you are
running more than a few backends concurrently, you really had better have a
real spinlock implementation if you want tolerable performance. And 1024
semaphores is a sizable fraction of the system-wide SysV semaphore limit
on many platforms. Therefore, reduce this setting's default value to 128
to make it less likely to cause out-of-semaphores problems.
Returning the direct result of bit arithmetic, in a macro intended to be
used in a boolean manner, can be problematic if the return value is
stored in a variable of type 'bool'. If bool is implemented using C99's
_Bool, that can lead to comparison failures if the variable is then
compared again with the expression (see ginStepRight() for an example
that fails), as _Bool forces the result to be 0/1. That happens in some
configurations of newer MSVC compilers. It's also problematic when
storing the result of such an expression in a narrower type.
Several gin macros have been declared in that style since gin's initial
commit in 8a3631f8d86.
There's a lot more macros like this, but this is the only one causing
regression test failures; and I don't want to commit and backpatch a
larger patch with lots of conflicts just before the next set of minor
releases.
Discussion: 20150811154237.GD17575@awork2.anarazel.de
Backpatch: All supported branches
Previously, we included <xlocale.h> only if necessary to get the definition
of type locale_t. According to notes in PGAC_TYPE_LOCALE_T, this is
important because on some versions of glibc that file supplies an
incompatible declaration of locale_t. (This info may be obsolete, because
on my RHEL6 box that seems to be the *only* definition of locale_t; but
there may still be glibc's in the wild for which it's a live concern.)
It turns out though that on FreeBSD and maybe other BSDen, you can get
locale_t from stdlib.h or locale.h but mbstowcs_l() and friends only from
<xlocale.h>. This was leaving us compiling calls to mbstowcs_l() and
friends with no visible prototype, which causes a warning and could
possibly cause actual trouble, since it's not declared to return int.
Hence, adjust the configure checks so that we'll include <xlocale.h>
either if it's necessary to get type locale_t or if it's necessary to
get a declaration of mbstowcs_l().
Report and patch by Aleksander Alekseev, somewhat whacked around by me.
Back-patch to all supported branches, since we have been using
mbstowcs_l() since 9.1.
Renaming a file using rename(2) is not guaranteed to be durable in face
of crashes; especially on filesystems like xfs and ext4 when mounted
with data=writeback. To be certain that a rename() atomically replaces
the previous file contents in the face of crashes and different
filesystems, one has to fsync the old filename, rename the file, fsync
the new filename, fsync the containing directory. This sequence is not
generally adhered to currently; which exposes us to data loss risks. To
avoid having to repeat this arduous sequence, introduce
durable_rename(), which wraps all that.
Also add durable_link_or_rename(). Several places use link() (with a
fallback to rename()) to rename a file, trying to avoid replacing the
target file out of paranoia. Some of those rename sequences need to be
durable as well. There seems little reason extend several copies of the
same logic, so centralize the link() callers.
This commit does not yet make use of the new functions; they're used in
a followup commit.
Author: Michael Paquier, Andres Freund
Discussion: 56583BDD.9060302@2ndquadrant.com
Backpatch: All supported branches
When decoding the old version of an UPDATE or DELETE change, and if that
tuple was bigger than MaxHeapTupleSize, we either Assert'ed out, or
failed in more subtle ways in non-assert builds. Normally individual
tuples aren't bigger than MaxHeapTupleSize, with big datums toasted.
But that's not the case for the old version of a tuple for logical
decoding; the replica identity is logged as one piece. With the default
replica identity btree limits that to small tuples, but that's not the
case for FULL.
Change the tuple buffer infrastructure to separate allocate over-large
tuples, instead of always going through the slab cache.
This unfortunately requires changing the ReorderBufferTupleBuf
definition, we need to store the allocated size someplace. To avoid
requiring output plugins to recompile, don't store HeapTupleHeaderData
directly after HeapTupleData, but point to it via t_data; that leaves
rooms for the allocated size. As there's no reason for an output plugin
to look at ReorderBufferTupleBuf->t_data.header, remove the field. It
was just a minor convenience having it directly accessible.
Reported-By: Adam Dratwiński
Discussion: CAKg6ypLd7773AOX4DiOGRwQk1TVOQKhNwjYiVjJnpq8Wo+i62Q@mail.gmail.com
Logical decoding's reorderbuffer keeps transactions in an LSN ordered
list for efficiency. To make that's efficiently possible upper-level
xids are forced to be logged before nested subtransaction xids. That
only works though if these records are all looked at: Unfortunately we
didn't do so for e.g. row level locks, which are otherwise uninteresting
for logical decoding.
This could lead to errors like:
"ERROR: subxact logged without previous toplevel record".
It's not sufficient to just look at row locking records, the xid could
appear first due to a lot of other types of records (which will trigger
the transaction to be marked logged with MarkCurrentTransactionIdLoggedIfAny).
So invent infrastructure to tell reorderbuffer about xids seen, when
they'd otherwise not pass through reorderbuffer.c.
Reported-By: Jarred Ward
Bug: #13844
Discussion: 20160105033249.1087.66040@wrigleys.postgresql.org
Backpatch: 9.4, where logical decoding was added
Reportedly, some compilers warn about tests like "c < 0" if c is unsigned,
and hence complain about the character range checks I added in commit
3bb3f42f3749d40b8d4de65871e8d828b18d4a45. This is a bit of a pain since
the regex library doesn't really want to assume that chr is unsigned.
However, since any such reconfiguration would involve manual edits of
regcustom.h anyway, we can put it on the shoulders of whoever wants to
do that to adjust this new range-checking macro correctly.
Per gripes from Coverity and Andres.
Previously, our regex code defined CHR_MAX as 0xfffffffe, which is a
bad choice because it is outside the range of type "celt" (int32).
Characters approaching that limit could lead to infinite loops in logic
such as "for (c = a; c <= b; c++)" where c is of type celt but the
range bounds are chr. Such loops will work safely only if CHR_MAX+1
is representable in celt, since c must advance to beyond b before the
loop will exit.
Fortunately, there seems no reason not to restrict CHR_MAX to 0x7ffffffe.
It's highly unlikely that Unicode will ever assign codes that high, and
none of our other backend encodings need characters beyond that either.
In addition to modifying the macro, we have to explicitly enforce character
range restrictions on the values of \u, \U, and \x escape sequences, else
the limit is trivially bypassed.
Also, the code for expanding case-independent character ranges in bracket
expressions had a potential integer overflow in its calculation of the
number of characters it could generate, which could lead to allocating too
small a character vector and then overwriting memory. An attacker with the
ability to supply arbitrary regex patterns could easily cause transient DOS
via server crashes, and the possibility for privilege escalation has not
been ruled out.
Quite aside from the integer-overflow problem, the range expansion code was
unnecessarily inefficient in that it always produced a result consisting of
individual characters, abandoning the knowledge that we had a range to
start with. If the input range is large, this requires excessive memory.
Change it so that the original range is reported as-is, and then we add on
any case-equivalent characters that are outside that range. With this
approach, we can bound the number of individual characters allowed without
sacrificing much. This patch allows at most 100000 individual characters,
which I believe to be more than the number of case pairs existing in
Unicode, so that the restriction will never be hit in practice.
It's still possible for range() to take awhile given a large character code
range, so also add statement-cancel detection to its loop. The downstream
function dovec() also lacked cancel detection, and could take a long time
given a large output from range().
Per fuzz testing by Greg Stark. Back-patch to all supported branches.
Security: CVE-2016-0773
All the other jsonb function descriptions refer to the arguments as being
"jsonb", but these two said "json". Make it consistent. Per bug #13905
from Petru Florin Mihancea.
No catversion bump --- we can't force one in the back branches, and this
isn't very critical anyway.
A scan for missed proisstrict markings in the core code turned up
these functions:
brin_summarize_new_values
pg_stat_reset_single_table_counters
pg_stat_reset_single_function_counters
pg_create_logical_replication_slot
pg_create_physical_replication_slot
pg_drop_replication_slot
The first three of these take OID, so a null argument will normally look
like a zero to them, resulting in "ERROR: could not open relation with OID
0" for brin_summarize_new_values, and no action for the pg_stat_reset_XXX
functions. The other three will dump core on a null argument, though this
is mitigated by the fact that they won't do so until after checking that
the caller is superuser or has rolreplication privilege.
In addition, the pg_logical_slot_get/peek[_binary]_changes family was
intentionally marked nonstrict, but failed to make nullness checks on all
the arguments; so again a null-pointer-dereference crash is possible but
only for superusers and rolreplication users.
Add the missing ARGISNULL checks to the latter functions, and mark the
former functions as strict in pg_proc. Make that change in the back
branches too, even though we can't force initdb there, just so that
installations initdb'd in future won't have the issue. Since none of these
bugs rise to the level of security issues (and indeed the pg_stat_reset_XXX
functions hardly misbehave at all), it seems sufficient to do this.
In addition, fix some order-of-operations oddities in the slot_get_changes
family, mostly cosmetic, but not the part that moves the function's last
few operations into the PG_TRY block. As it stood, there was significant
risk for an error to exit without clearing historical information from
the system caches.
The slot_get_changes bugs go back to 9.4 where that code was introduced.
Back-patch appropriate subsets of the pg_proc changes into all active
branches, as well.
This is necessary so that REASSIGN OWNED does the right thing with
composite types, to wit, that it also alters ownership of the type's
pg_class entry -- previously, the pg_class entry remained owned by the
original user, which caused later other failures such as the new owner's
inability to use ALTER TYPE to rename an attribute of the affected
composite. Also, if the original owner is later dropped, the pg_class
entry becomes owned by a non-existant user which is bogus.
To fix, create a new routine AlterTypeOwner_oid which knows whether to
pass the request to ATExecChangeOwner or deal with it directly, and use
that in shdepReassignOwner rather than calling AlterTypeOwnerInternal
directly. AlterTypeOwnerInternal is now simpler in that it only
modifies the pg_type entry and recurses to handle a possible array type;
higher-level tasks are handled by either AlterTypeOwner directly or
AlterTypeOwner_oid.
I took the opportunity to add a few more objects to the test rig for
REASSIGN OWNED, so that more cases are exercised. Additional ones could
be added for superuser-only-ownable objects (such as FDWs and event
triggers) but I didn't want to push my luck by adding a new superuser to
the tests on a backpatchable bug fix.
Per bug #13666 reported by Chris Pacejo.
This is a backpatch of commit 756e7b4c9db1 to branches 9.1 -- 9.4.
It emerges that libreadline doesn't notice terminal window size change
events unless they occur while collecting input. This is easy to stumble
over if you resize the window while using a pager to look at query output,
but it can be demonstrated without any pager involvement. The symptom is
that queries exceeding one line are misdisplayed during subsequent input
cycles, because libreadline has the wrong idea of the screen dimensions.
The safest, simplest way to fix this is to call rl_reset_screen_size()
just before calling readline(). That causes an extra ioctl(TIOCGWINSZ)
for every command; but since it only happens when reading from a tty, the
performance impact should be negligible. A more valid objection is that
this still leaves a tiny window during entry to readline() wherein delivery
of SIGWINCH will be missed; but the practical consequences of that are
probably negligible. In any case, there doesn't seem to be any good way to
avoid the race, since readline exposes no functions that seem safe to call
from a generic signal handler --- rl_reset_screen_size() certainly isn't.
It turns out that we also need an explicit rl_initialize() call, else
rl_reset_screen_size() dumps core when called before the first readline()
call.
rl_reset_screen_size() is not present in old versions of libreadline,
so we need a configure test for that. (rl_initialize() is present at
least back to readline 4.0, so we won't bother with a test for it.)
We would need a configure test anyway since libedit's emulation of
libreadline doesn't currently include such a function. Fortunately,
libedit seems not to have any corresponding bug.
Merlin Moncure, adjusted a bit by me
More fuzz testing by Andreas Seltenreich exposed that the planner did not
cope well with chains of lateral references. If relation X references Y
laterally, and Y references Z laterally, then we will have to scan X on the
inside of a nestloop with Z, so for all intents and purposes X is laterally
dependent on Z too. The planner did not understand this and would generate
intermediate joins that could not be used. While that was usually harmless
except for wasting some planning cycles, under the right circumstances it
would lead to "failed to build any N-way joins" or "could not devise a
query plan" planner failures.
To fix that, convert the existing per-relation lateral_relids and
lateral_referencers relid sets into their transitive closures; that is,
they now show all relations on which a rel is directly or indirectly
laterally dependent. This not only fixes the chained-reference problem
but allows some of the relevant tests to be made substantially simpler
and faster, since they can be reduced to simple bitmap manipulations
instead of searches of the LateralJoinInfo list.
Also, when a PlaceHolderVar that is due to be evaluated at a join contains
lateral references, we should treat those references as indirect lateral
dependencies of each of the join's base relations. This prevents us from
trying to join any individual base relations to the lateral reference
source before the join is formed, which again cannot work.
Andreas' testing also exposed another oversight in the "dangerous
PlaceHolderVar" test added in commit 85e5e222b1dd02f1. Simply rejecting
unsafe join paths in joinpath.c is insufficient, because in some cases
we will end up rejecting *all* possible paths for a particular join, again
leading to "could not devise a query plan" failures. The restriction has
to be known also to join_is_legal and its cohort functions, so that they
will not select a join for which that will happen. I chose to move the
supporting logic into joinrels.c where the latter functions are.
Back-patch to 9.3 where LATERAL support was introduced.
At the end of crash recovery, unlogged relations are reset to the empty
state, using their init fork as the template. The init fork is copied to
the main fork without going through shared buffers. Unfortunately WAL
replay so far has not necessarily flushed writes from shared buffers to
disk at that point. In normal crash recovery, and before the
introduction of 'fast promotions' in fd4ced523 / 9.3, the
END_OF_RECOVERY checkpoint flushes the buffers out in time. But with
fast promotions that's not the case anymore.
To fix, force WAL writes targeting the init fork to be flushed
immediately (using the new FlushOneBuffer() function). In 9.5+ that
flush can centrally be triggered from the code dealing with restoring
full page writes (XLogReadBufferForRedoExtended), in earlier releases
that responsibility is in the hands of XLOG_HEAP_NEWPAGE's replay
function.
Backpatch to 9.1, even if this currently is only known to trigger in
9.3+. Flushing earlier is more robust, and it is advantageous to keep
the branches similar.
Typical symptoms of this bug are errors like
'ERROR: index "..." contains unexpected zero page at block 0'
shortly after promoting a node.
Reported-By: Thom Brown
Author: Andres Freund and Michael Paquier
Discussion: 20150326175024.GJ451@alap3.anarazel.de
Backpatch: 9.1-
While convincing myself that commit 7e19db0c09719d79 would solve both of
the problems recently reported by Andreas Seltenreich, I realized that
add_paths_to_joinrel's handling of LATERAL restrictions could be made
noticeably simpler and faster if we were to retain the minimum possible
parameterization for each joinrel (that is, the set of relids supplying
unsatisfied lateral references in it). We already retain that for
baserels, in RelOptInfo.lateral_relids, so we can use that field for
joinrels too.
This is a back-port of commit edca44b1525b3d591263d032dc4fe500ea771e0e.
I originally intended not to back-patch that, but additional hacking
in this area turns out to be needed, making it necessary not optional
to compute lateral_relids for joinrels. In preparation for those fixes,
sync the relevant code with HEAD as much as practical. (I did not risk
rearranging fields of RelOptInfo in released branches, however.)
It was possible for the planner to decide to join a LATERAL subquery to
the outer side of an outer join before the outer join itself is completed.
Normally that's fine because of the associativity rules, but it doesn't
work if the subquery contains a lateral reference to the inner side of the
outer join. In such a situation the outer join *must* be done first.
join_is_legal() missed this consideration and would allow the join to be
attempted, but the actual path-building code correctly decided that no
valid join path could be made, sometimes leading to planner errors such as
"failed to build any N-way joins".
Per report from Andreas Seltenreich. Back-patch to 9.3 where LATERAL
support was added.
The POSIX standard for tar headers requires archive member sizes to be
printed in octal with at most 11 digits, limiting the representable file
size to 8GB. However, GNU tar and apparently most other modern tars
support a convention in which oversized values can be stored in base-256,
allowing any practical file to be a tar member. Adopt this convention
to remove two limitations:
* pg_dump with -Ft output format failed if the contents of any one table
exceeded 8GB.
* pg_basebackup failed if the data directory contained any file exceeding
8GB. (This would be a fatal problem for installations configured with a
table segment size of 8GB or more, and it has also been seen to fail when
large core dump files exist in the data directory.)
File sizes under 8GB are still printed in octal, so that no compatibility
issues are created except in cases that would have failed entirely before.
In addition, this patch fixes several bugs in the same area:
* In 9.3 and later, we'd defined tarCreateHeader's file-size argument as
size_t, which meant that on 32-bit machines it would write a corrupt tar
header for file sizes between 4GB and 8GB, even though no error was raised.
This broke both "pg_dump -Ft" and pg_basebackup for such cases.
* pg_restore from a tar archive would fail on tables of size between 4GB
and 8GB, on machines where either "size_t" or "unsigned long" is 32 bits.
This happened even with an archive file not affected by the previous bug.
* pg_basebackup would fail if there were files of size between 4GB and 8GB,
even on 64-bit machines.
* In 9.3 and later, "pg_basebackup -Ft" failed entirely, for any file size,
on 64-bit big-endian machines.
In view of these potential data-loss bugs, back-patch to all supported
branches, even though removal of the documented 8GB limit might otherwise
be considered a new feature rather than a bug fix.
Commit bda76c1c8cfb1d11751ba6be88f0242850481733 caused both plus and
minus infinity to be rendered as "infinity", which is not only wrong
but inconsistent with the pre-9.4 behavior of to_json(). Fix that by
duplicating the coding in date_out/timestamp_out/timestamptz_out more
closely. Per bug #13687 from Stepan Perlov. Back-patch to 9.4, like
the previous commit.
In passing, also re-pgindent json.c, since it had gotten a bit messed up by
recent patches (and I was already annoyed by indentation-related problems
in back-patching this fix ...)
This code previously counted the number of NFA states it created, and
complained if a limit was exceeded, so as to prevent bizarre regex patterns
from consuming unreasonable time or memory. That's fine as far as it went,
but the code paid no attention to how many arcs linked those states. Since
regexes can be contrived that have O(N) states but will need O(N^2) arcs
after fixempties() processing, it was still possible to blow out memory,
and take a long time doing it too. To fix, modify the bookkeeping to count
space used by both states and arcs.
I did not bother with including the "color map" in the accounting; it
can only grow to a few megabytes, which is not a lot in comparison to
what we're allowing for states+arcs (about 150MB on 64-bit machines
or half that on 32-bit machines).
Looking at some of the larger real-world regexes captured in the Tcl
regression test suite suggests that the most that is likely to be needed
for regexes found in the wild is under 10MB, so I believe that the current
limit has enough headroom to make it okay to keep it as a hard-wired limit.
In connection with this, redefine REG_ETOOBIG as meaning "regular
expression is too complex"; the previous wording of "nfa has too many
states" was already somewhat inapropos because of the error code's use
for stack depth overrun, and it was not very user-friendly either.
Back-patch to all supported branches.
Change the singly-linked in-arc and out-arc lists to be doubly-linked,
so that arc deletion is constant time rather than having worst-case time
proportional to the number of other arcs on the connected states.
Modify the bulk arc transfer operations copyins(), copyouts(), moveins(),
moveouts() so that they use a sort-and-merge algorithm whenever there's
more than a small number of arcs to be copied or moved. The previous
method is O(N^2) in the number of arcs involved, because it performs
duplicate checking independently for each copied arc. The new method may
change the ordering of existing arcs for the destination state, but nothing
really cares about that.
Provide another bulk arc copying method mergeins(), which is unused as
of this commit but is needed for the next one. It basically is like
copyins(), but the source arcs might not all come from the same state.
Replace the O(N^2) bubble-sort algorithm used in carcsort() with a qsort()
call.
These changes greatly improve the performance of regex compilation for
large or complex regexes, at the cost of extra space for arc storage during
compilation. The original tradeoff was probably fine when it was made, but
now we care more about speed and less about memory consumption.
Back-patch to all supported branches.
Postmaster child processes that aren't supposed to be attached to shared
memory were not bothering to close the shared memory mapping handle they
inherit from the postmaster process. That's mostly harmless, since the
handle vanishes anyway when the child process exits -- but the syslogger
process, if used, doesn't get killed and restarted during recovery from a
backend crash. That meant that Windows doesn't see the shared memory
mapping as becoming free, so it doesn't delete it and the postmaster is
unable to create a new one, resulting in failure to recover from crashes
whenever logging_collector is turned on.
Per report from Dmitry Vasilyev. It's a bit astonishing that we'd not
figured this out long ago, since it's been broken from the very beginnings
of out native Windows support; probably some previously-unexplained trouble
reports trace to this.
A secondary problem is that on Cygwin (perhaps only in older versions?),
exec() may not detach from the shared memory segment after all, in which
case these child processes did remain attached to shared memory, posing
the risk of an unexpected shared memory clobber if they went off the rails
somehow. That may be a long-gone bug, but we can deal with it now if it's
still live, by detaching within the infrastructure introduced here to deal
with closing the handle.
Back-patch to all supported branches.
Tom Lane and Amit Kapila
The postmaster now checks every minute or so (worst case, at most two
minutes) that postmaster.pid is still there and still contains its own PID.
If not, it performs an immediate shutdown, as though it had received
SIGQUIT.
The original goal behind this change was to ensure that failed buildfarm
runs would get fully cleaned up, even if the test scripts had left a
postmaster running, which is not an infrequent occurrence. When the
buildfarm script removes a test postmaster's $PGDATA directory, its next
check on postmaster.pid will fail and cause it to exit. Previously, manual
intervention was often needed to get rid of such orphaned postmasters,
since they'd block new test postmasters from obtaining the expected socket
address.
However, by checking postmaster.pid and not something else, we can provide
additional robustness: manual removal of postmaster.pid is a frequent DBA
mistake, and now we can at least limit the damage that will ensue if a new
postmaster is started while the old one is still alive.
Back-patch to all supported branches, since we won't get the desired
improvement in buildfarm reliability otherwise.
Some of the functions in regex compilation and execution recurse, and
therefore could in principle be driven to stack overflow. The Tcl crew
has seen this happen in practice in duptraverse(), though their fix was
to put in a hard-wired limit on the number of recursive levels, which is
not too appetizing --- fortunately, we have enough infrastructure to check
the actually available stack. Greg Stark has also seen it in other places
while fuzz testing on a machine with limited stack space. Let's put guards
in to prevent crashes in all these places.
Since the regex code would leak memory if we simply threw elog(ERROR),
we have to introduce an API that checks for stack depth without throwing
such an error. Fortunately that's not difficult.
After an internal failure in shortest() or longest() while pinning down the
exact location of a match, find() forgot to free the DFA structure before
returning. This is pretty unlikely to occur, since we just successfully
ran the "search" variant of the DFA; but it could happen, and it would
result in a session-lifespan memory leak since this code uses malloc()
directly. Problem seems to have been aboriginal in Spencer's library,
so back-patch all the way.
In passing, correct a thinko in a comment I added awhile back about the
meaning of the "ntree" field.
I happened across these issues while comparing our code to Tcl's version
of the library.
This commit makes postmaster forcibly remove the files signaling
a standby promotion request. Otherwise, the existence of those files
can trigger a promotion too early, whether a user wants that or not.
This removal of files is usually unnecessary because they can exist
only during a few moments during a standby promotion. However
there is a race condition: if pg_ctl promote is executed and creates
the files during a promotion, the files can stay around even after
the server is brought up to new master. Then, if new standby starts
by using the backup taken from that master, the files can exist
at the server startup and should be removed in order to avoid
an unexpected promotion.
Back-patch to 9.1 where promote signal file was introduced.
Problem reported by Feike Steenbergen.
Original patch by Michael Paquier, modified by me.
Discussion: 20150528100705.4686.91426@wrigleys.postgresql.org
Formerly, we treated only portals created in the current subtransaction as
having failed during subtransaction abort. However, if the error occurred
while running a portal created in an outer subtransaction (ie, a cursor
declared before the last savepoint), that has to be considered broken too.
To allow reliable detection of which ones those are, add a bookkeeping
field to struct Portal that tracks the innermost subtransaction in which
each portal has actually been executed. (Without this, we'd end up
failing portals containing functions that had called the subtransaction,
thereby breaking plpgsql exception blocks completely.)
In addition, when we fail an outer-subtransaction Portal, transfer its
resources into the subtransaction's resource owner, so that they're
released early in cleanup of the subxact. This fixes a problem reported by
Jim Nasby in which a function executed in an outer-subtransaction cursor
could cause an Assert failure or crash by referencing a relation created
within the inner subtransaction.
The proximate cause of the Assert failure is that AtEOSubXact_RelationCache
assumed it could blow away a relcache entry without first checking that the
entry had zero refcount. That was a bad idea on its own terms, so add such
a check there, and to the similar coding in AtEOXact_RelationCache. This
provides an independent safety measure in case there are still ways to
provoke the situation despite the Portal-level changes.
This has been broken since subtransactions were invented, so back-patch
to all supported branches.
Tom Lane and Michael Paquier
It's only necessary to fix this in 9.4; later versions don't have this
code (because we ripped out Alpha support entirely), while earlier ones
aren't actually using pg_read_barrier() anywhere.
Per rather belated report from Christoph Berg.
On recent AIX it's necessary to configure gcc to use the native assembler
(because the GNU assembler hasn't been updated to handle AIX 6+). This
caused PG builds to fail with assembler syntax errors, because we'd try
to compile s_lock.h's gcc asm fragment for PPC, and that assembly code
relied on GNU-style local labels. We can't substitute normal labels
because it would fail in any file containing more than one inlined use of
tas(). Fortunately, that code is stable enough, and the PPC ISA is simple
enough, that it doesn't seem like too much of a maintenance burden to just
hand-code the branch offsets, removing the need for any labels.
Note that the AIX assembler only accepts "$" for the location counter
pseudo-symbol. The usual GNU convention is "."; but it appears that all
versions of gas for PPC also accept "$", so in theory this patch will not
break any other PPC platforms.
This has been reported by a few people, but Steve Underwood gets the credit
for being the first to pursue the problem far enough to understand why it
was failing. Thanks also to Noah Misch for additional testing.
Formerly, this function would always return "true" for an appendrel child
relation, because it would think that the appendrel parent was a potential
join target for the child. In principle that should only lead to some
inefficiency in planning, but fuzz testing by Andreas Seltenreich disclosed
that it could lead to "could not find pathkey item to sort" planner errors
in odd corner cases. Specifically, we would think that all columns of a
child table's multicolumn index were interesting pathkeys, causing us to
generate a MergeAppend path that sorts by all the columns. However, if any
of those columns weren't actually used above the level of the appendrel,
they would not get added to that rel's targetlist, which would result in
being unable to resolve the MergeAppend's sort keys against its targetlist
during createplan.c.
Backpatch to 9.3. In older versions, columns of an appendrel get added
to its targetlist even if they're not mentioned above the scan level,
so that the failure doesn't occur. It might be worth back-patching this
fix to older versions anyway, but I'll refrain for the moment.
The tuplesort/tuplestore memory management logic assumed that the chunk
allocation overhead for its memtuples array could not increase when
increasing the array size. This is and always was true for tuplesort,
but we (I, I think) blindly copied that logic into tuplestore.c without
noticing that the assumption failed to hold for the much smaller array
elements used by tuplestore. Given rather small work_mem, this could
result in an improper complaint about "unexpected out-of-memory situation",
as reported by Brent DeSpain in bug #13530.
The easiest way to fix this is just to increase tuplestore's initial
array size so that the assumption holds. Rather than relying on magic
constants, though, let's export a #define from aset.c that represents
the safe allocation threshold, and make tuplestore's calculation depend
on that.
Do the same in tuplesort.c to keep the logic looking parallel, even though
tuplesort.c isn't actually at risk at present. This will keep us from
breaking it if we ever muck with the allocation parameters in aset.c.
Back-patch to all supported versions. The error message doesn't occur
pre-9.3, not so much because the problem can't happen as because the
pre-9.3 tuplestore code neglected to check for it. (The chance of
trouble is a great deal larger as of 9.3, though, due to changes in the
array-size-increasing strategy.) However, allowing LACKMEM() to become
true unexpectedly could still result in less-than-desirable behavior,
so let's patch it all the way back.
Commit c9b0cbe98bd783e24a8c4d8d8ac472a494b81292 accidentally broke the
order of operations during postmaster shutdown: it resulted in removing
the per-socket lockfiles after, not before, postmaster.pid. This creates
a race-condition hazard for a new postmaster that's started immediately
after observing that postmaster.pid has disappeared; if it sees the
socket lockfile still present, it will quite properly refuse to start.
This error appears to be the explanation for at least some of the
intermittent buildfarm failures we've seen in the pg_upgrade test.
Another problem, which has been there all along, is that the postmaster
has never bothered to close() its listen sockets, but has just allowed them
to close at process death. This creates a different race condition for an
incoming postmaster: it might be unable to bind to the desired listen
address because the old postmaster is still incumbent. This might explain
some odd failures we've seen in the past, too. (Note: this is not related
to the fact that individual backends don't close their client communication
sockets. That behavior is intentional and is not changed by this patch.)
Fix by adding an on_proc_exit function that closes the postmaster's ports
explicitly, and (in 9.3 and up) reshuffling the responsibility for where
to unlink the Unix socket files. Lock file unlinking can stay where it
is, but teach it to unlink the lock files in reverse order of creation.
SP-GiST initialized an all-zeros page at vacuum, but that was not
WAL-logged, which is not safe. You might get a torn page write, when it gets
flushed to disk, and end-up with a half-initialized index page. To fix,
leave it in the all-zeros state, and add it to the FSM. It will be
initialized when reused. Also don't set the page-deleted flag when recycling
an empty page. That was also not WAL-logged, and a torn write of that would
cause the page to have an invalid checksum.
Backpatch to 9.2, where SP-GiST indexes were added.
