These bugs can cause data loss on standbys started with hot_standby=on at
the moment they start to accept read only queries, by marking committed
transactions as uncommited. The likelihood of such corruptions is small
unless the primary has a high transaction rate.
5a031a5556 fixed bugs in HS's startup logic
by maintaining less state until at least STANDBY_SNAPSHOT_PENDING state
was reached, missing the fact that both clog and subtrans are written to
before that. This only failed to fail in common cases because the usage
of ExtendCLOG in procarray.c was superflous since clog extensions are
actually WAL logged.
f44eedc3f0f347a856eea8590730769125964597/I then tried to fix the missing
extensions of pg_subtrans due to the former commit's changes - which are
not WAL logged - by performing the extensions when switching to a state
> STANDBY_INITIALIZED and not performing xid assignments before that -
again missing the fact that ExtendCLOG is unneccessary - but screwed up
twice: Once because latestObservedXid wasn't updated anymore in that
state due to the earlier commit and once by having an off-by-one error in
the loop performing extensions. This means that whenever a
CLOG_XACTS_PER_PAGE (32768 with default settings) boundary was crossed
between the start of the checkpoint recovery started from and the first
xl_running_xact record old transactions commit bits in pg_clog could be
overwritten if they started and committed in that window.
Fix this mess by not performing ExtendCLOG() in HS at all anymore since
it's unneeded and evidently dangerous and by performing subtrans
extensions even before reaching STANDBY_SNAPSHOT_PENDING.
Analysis and patch by Andres Freund. Reported by Christophe Pettus.
Backpatch down to 9.0, like the previous commit that caused this.
A couple of places that should have been iterating over WORDS_PER_CHUNK
words were iterating over WORDS_PER_PAGE words instead. This thinko
accidentally failed to fail, because (at least on common architectures
with default BLCKSZ) WORDS_PER_CHUNK is a bit less than WORDS_PER_PAGE,
and the extra words being looked at were always zero so nothing happened.
Still, it's a bug waiting to happen if anybody ever fools with the
parameters affecting TIDBitmap sizes, and it's a small waste of cycles
too. So back-patch to all active branches.
Etsuro Fujita
If a page is deleted, and reused for something else, just as a search is
following a rightlink to it from its left sibling, the search would continue
scanning whatever the new contents of the page are. That could lead to
incorrect query results, or even something more curious if the page is
reused for a different kind of a page.
To fix, modify the search algorithm to lock the next page before releasing
the previous one, and refrain from deleting pages from the leftmost branch
of the tree.
Add a new Concurrency section to the README, explaining why this works.
There is a lot more one could say about concurrency in GIN, but that's for
another patch.
Backpatch to all supported versions.
This change prevents us from doing inappropriate subquery flattening in
cases such as dangerous functions hidden inside a sub-SELECT in the
targetlist of another sub-SELECT. That could result in unexpected behavior
due to multiple evaluations of a volatile function, as in a recent
complaint from Etienne Dube. It's been questionable from the very
beginning whether these functions should look into subqueries (as noted in
their comments), and this case seems to provide proof that they should.
Because the new code only descends into SubLinks, not SubPlans or
InitPlans, the change only affects the planner's behavior during
prepjointree processing and not later on --- for example, you can still get
it to use a volatile function in an indexqual if you wrap the function in
(SELECT ...). That's a historical behavior, for sure, but it's reasonable
given that the executor's evaluation rules for subplans don't depend on
whether there are volatile functions inside them. In any case, we need to
constrain the behavioral change as narrowly as we can to make this
reasonable to back-patch.
Back-patch commits 8e68816cc2 and
8dace66e07 into the stable branches.
Buildfarm testing revealed no great portability surprises, and it
seems useful to have this robustness improvement in all branches.
Before jamming a desired targetlist into a plan node, one really ought to
make sure the plan node can handle projections, and insert a buffering
Result plan node if not. planagg.c forgot to do this, which is a hangover
from the days when it only dealt with IndexScan plan types. MergeAppend
doesn't project though, not to mention that it gets unhappy if you remove
its possibly-resjunk sort columns. The code accidentally failed to fail
for cases in which the min/max argument was a simple Var, because the new
targetlist would be equivalent to the original "flat" tlist anyway.
For any more complex case, it's been broken since 9.1 where we introduced
the ability to optimize min/max using MergeAppend, as reported by Raphael
Bauduin. Fix by duplicating the logic from grouping_planner that decides
whether we need a Result node.
In 9.2 and 9.1, this requires back-porting the tlist_same_exprs() function
introduced in commit 4387cf956b, else we'd
uselessly add a Result node in cases that worked before. It's rather
tempting to back-patch that whole commit so that we can avoid extra Result
nodes in mainline cases too; but I'll refrain, since that code hasn't
really seen all that much field testing yet.
Insertion of default arguments doesn't work for window functions, which is
likely to cause a crash at runtime if the implementation code doesn't check
the number of actual arguments carefully. It doesn't seem worth working
harder than this for pre-9.2 branches.
For rather inscrutable reasons, SQL:2008 disallows copying-and-modifying a
window definition that has any explicit framing clause. The error message
we gave for this only made sense if the referencing window definition
itself contains an explicit framing clause, which it might well not.
Moreover, in the context of an OVER clause it's not exactly obvious that
"OVER (windowname)" implies copy-and-modify while "OVER windowname" does
not. This has led to multiple complaints, eg bug #5199 from Iliya
Krapchatov. Change to a hopefully more intelligible error message, and
in the case where we have just "OVER (windowname)", add a HINT suggesting
that omitting the parentheses will fix it. Also improve the related
documentation. Back-patch to all supported branches.
Callers expect that they only have to set the right resource owner when
creating a BufFile, not during subsequent operations on it. While we could
insist this be fixed at the caller level, it seems more sensible for the
BufFile to take care of it. Without this, some temp files belonging to
a BufFile can go away too soon, eg at the end of a subtransaction,
leading to errors or crashes.
Reported and fixed by Andres Freund. Back-patch to all active branches.
Formerly, when using a SQL-spec timezone setting with a fixed GMT offset
(called a "brute force" timezone in the code), the session_timezone
variable was not updated to match the nominal timezone; rather, all code
was expected to ignore session_timezone if HasCTZSet was true. This is
of course obviously fragile, though a search of the code finds only
timeofday() failing to honor the rule. A bigger problem was that
DetermineTimeZoneOffset() supposed that if its pg_tz parameter was
pointer-equal to session_timezone, then HasCTZSet should override the
parameter. This would cause datetime input containing an explicit zone
name to be treated as referencing the brute-force zone instead, if the
zone name happened to match the session timezone that had prevailed
before installing the brute-force zone setting (as reported in bug #8572).
The same malady could affect AT TIME ZONE operators.
To fix, set up session_timezone so that it matches the brute-force zone
specification, which we can do using the POSIX timezone definition syntax
"<abbrev>offset", and get rid of the bogus lookaside check in
DetermineTimeZoneOffset(). Aside from fixing the erroneous behavior in
datetime parsing and AT TIME ZONE, this will cause the timeofday() function
to print its result in the user-requested time zone rather than some
previously-set zone. It might also affect results in third-party
extensions, if there are any that make use of session_timezone without
considering HasCTZSet, but in all cases the new behavior should be saner
than before.
Back-patch to all supported branches.
The C and POSIX standards state that strncpy's behavior is undefined when
source and destination areas overlap. While it remains dubious whether any
implementations really misbehave when the pointers are exactly equal, some
platforms are now starting to force the issue by complaining when an
undefined call occurs. (In particular OS X 10.9 has been seen to dump core
here, though the exact set of circumstances needed to trigger that remain
elusive. Similar behavior can be expected to be optional on Linux and
other platforms in the near future.) So tweak the code to explicitly do
nothing when nothing need be done.
Back-patch to all active branches. In HEAD, this also lets us get rid of
an exception in valgrind.supp.
Per discussion of a report from Matthias Schmitt.
1. In heap_hot_search_buffer(), the PredicateLockTuple() call is passed
wrong offset number. heapTuple->t_self is set to the tid of the first
tuple in the chain that's visited, not the one actually being read.
2. CheckForSerializableConflictIn() uses the tuple's t_ctid field
instead of t_self to check for exiting predicate locks on the tuple. If
the tuple was updated, but the updater rolled back, t_ctid points to the
aborted dead tuple.
Reported by Hannu Krosing. Backpatch to 9.1.
If a tuple was frozen while its predicate locks mattered,
read-write dependencies could be missed, resulting in failure to
detect conflicts which could lead to anomalies in committed
serializable transactions.
This field was added to the tag when we still thought that it was
necessary to carry locks forward to a new version of an updated
row. That was later proven to be unnecessary, which allowed
simplification of the code, but elimination of xmin from the tag
was missed at the time.
Per report and analysis by Heikki Linnakangas.
Backpatch to 9.1.
lo_open registers the currently active snapshot, and checks if the
large object exists after that. Normally, snapshots registered by lo_open
are unregistered at end of transaction when the lo descriptor is closed, but
if we error out before the lo descriptor is added to the list of open
descriptors, it is leaked. Fix by moving the snapshot registration to after
checking if the large object exists.
Reported by Pavel Stehule. Backpatch to 8.4. The snapshot registration
system was introduced in 8.4, so prior versions are not affected (and not
supported, anyway).
This is a backpatch of commits d942f9d9, 82b01026, and 6697aa2bc, back
to release 9.1 where we introduced extensions which make heavy use of
the PGXS infrastructure.
There is a rare race condition, when a transaction that inserted a tuple
aborts while vacuum is processing the page containing the inserted tuple.
Vacuum prunes the page first, which normally removes any dead tuples, but
if the inserting transaction aborts right after that, the loop after
pruning will see a dead tuple and remove it instead. That's OK, but if the
page is on a table with no indexes, and the page becomes completely empty
after removing the dead tuple (or tuples) on it, it will be immediately
marked as all-visible. That's OK, but the sanity check in vacuum would
throw a warning because it thinks that the page contains dead tuples and
was nevertheless marked as all-visible, even though it just vacuumed away
the dead tuples and so it doesn't actually contain any.
Spotted this while reading the code. It's difficult to hit the race
condition otherwise, but can be done by putting a breakpoint after the
heap_page_prune() call.
Backpatch all the way to 8.4, where this code first appeared.
Though @libdir@ almost always matches @abs_builddir@ in this context,
the test could only fail if they differed. Back-patch to 9.1, where the
test was introduced.
Hamid Quddus Akhtar
In libpq, we set up and pass to OpenSSL callback routines to handle
locking. When we run out of SSL connections, we try to clean things
up by de-registering the hooks. Unfortunately, we had a few calls
into the OpenSSL library after these hooks were de-registered during
SSL cleanup which lead to deadlocking. This moves the thread callback
cleanup to be after all SSL-cleanup related OpenSSL library calls.
I've been unable to reproduce the deadlock with this fix.
In passing, also move the close_SSL call to be after unlocking our
ssl_config mutex when in a failure state. While it looks pretty
unlikely to be an issue, it could have resulted in deadlocks if we
ended up in this code path due to something other than SSL_new
failing. Thanks to Heikki for pointing this out.
Back-patch to all supported versions; note that the close_SSL issue
only goes back to 9.0, so that hunk isn't included in the 8.4 patch.
Initially found and reported by Vesa-Matti J Kari; many thanks to
both Heikki and Andres for their help running down the specific
issue and reviewing the patch.
Once the administrator has called for an immediate shutdown or a backend
crash has triggered a reinitialization, no mere SIGINT or SIGTERM should
change that course. Such derailment remains possible when the signal
arrives before quickdie() blocks signals. That being a narrow race
affecting most PostgreSQL signal handlers in some way, leave it for
another patch. Back-patch this to all supported versions.
The previous coding attempted to activate all the GUC settings specified
in SET clauses, so that the function validator could operate in the GUC
environment expected by the function body. However, this is problematic
when restoring a dump, since the SET clauses might refer to database
objects that don't exist yet. We already have the parameter
check_function_bodies that's meant to prevent forward references in
function definitions from breaking dumps, so let's change CREATE FUNCTION
to not install the SET values if check_function_bodies is off.
Authors of function validators were already advised not to make any
"context sensitive" checks when check_function_bodies is off, if indeed
they're checking anything at all in that mode. But extend the
documentation to point out the GUC issue in particular.
(Note that we still check the SET clauses to some extent; the behavior
with !check_function_bodies is now approximately equivalent to what ALTER
DATABASE/ROLE have been doing for awhile with context-dependent GUCs.)
This problem can be demonstrated in all active branches, so back-patch
all the way.
This change will only apply to mingw compilers, and has been found
necessary by late versions of the mingw-w64 compiler. It's the same as
what is done elsewhere for the Microsoft compilers.
Backpatch of commit 73838b5251.
Problem reported by Michael Cronenworth, although not his patch.
It's possible that inlining of SQL functions (or perhaps other changes?)
has exposed typmod information not known at parse time. In such cases,
Vars generated by query_planner might have valid typmod values while the
original grouping columns only have typmod -1. This isn't a semantic
problem since the behavior of grouping only depends on type not typmod,
but it breaks locate_grouping_columns' use of tlist_member to locate the
matching entry in query_planner's result tlist.
We can fix this without an excessive amount of new code or complexity by
relying on the fact that locate_grouping_columns only gets called when
make_subplanTargetList has set need_tlist_eval == false, and that can only
happen if all the grouping columns are simple Vars. Therefore we only need
to search the sub_tlist for a matching Var, and we can reasonably define a
"match" as being a match of the Var identity fields
varno/varattno/varlevelsup. The code still Asserts that vartype matches,
but ignores vartypmod.
Per bug #8393 from Evan Martin. The added regression test case is
basically the same as his example. This has been broken for a very long
time, so back-patch to all supported branches.
With this optimization flag enabled, recent versions of gcc can generate
incorrect code that assumes variable-length arrays (such as oidvector)
are actually fixed-length because they're embedded in some larger struct.
The known instance of this problem was fixed in 9.2 and up by commit
8137f2c323 and followon work, which hides
actually-variable-length catalog fields from the compiler altogether.
And we plan to gradually convert variable-length fields to official
"flexible array member" notation over time, which should prevent this type
of bug from reappearing as gcc gets smarter. We're not going to try to
back-port those changes into older branches, though, so apply this
band-aid instead.
Andres Freund
The C99 and POSIX standards require strtod() to accept all these spellings
(case-insensitively): "inf", "+inf", "-inf", "infinity", "+infinity",
"-infinity". However, pre-C99 systems might accept only some or none of
these, and apparently Windows still doesn't accept "inf". To avoid
surprising cross-platform behavioral differences, manually check for each
of these spellings if strtod() fails. We were previously handling just
"infinity" and "-infinity" that way, but since C99 is most of the world
now, it seems likely that applications are expecting all these spellings
to work.
Per bug #8355 from Basil Peace. It turns out this fix won't actually
resolve his problem, because Python isn't being this careful; but that
doesn't mean we shouldn't be.
If a tuple is locked but not updated by a concurrent transaction,
HeapTupleSatisfiesDirty would return that transaction's Xid in xmax,
causing callers to wait on it, when it is not necessary (in fact, if the
other transaction had used a multixact instead of a plain Xid to mark
the tuple, HeapTupleSatisfiesDirty would have behave differently and
*not* returned the Xmax).
This bug was introduced in commit 3f7fbf85dc, dated December 1998,
so it's almost 15 years old now. However, it's hard to see this
misbehave, because before we had NOWAIT the only consequence of this is
that transactions would wait for slightly more time than necessary; so
it's not surprising that this hasn't been reported yet.
Craig Ringer and Andres Freund
We should really be reporting a useful error along with returning
a valid return code if pthread_mutex_lock() throws an error for
some reason. Add that and back-patch to 9.0 as the prior patch.
Pointed out by Alvaro Herrera
I've been working with Nick Phillips on an issue he ran into when
trying to use threads with SSL client certificates. As it turns out,
the call in initialize_SSL() to SSL_CTX_use_certificate_chain_file()
will modify our SSL_context without any protection from other threads
also calling that function or being at some other point and trying to
read from SSL_context.
To protect against this, I've written up the attached (based on an
initial patch from Nick and much subsequent discussion) which puts
locks around SSL_CTX_use_certificate_chain_file() and all of the other
users of SSL_context which weren't already protected.
Nick Phillips, much reworked by Stephen Frost
Back-patch to 9.0 where we started loading the cert directly instead of
using a callback.
Heikki Linnakangas