| Age | Commit message (Collapse) | Author |
|---|
|
Commit 85e5e222b1dd02f135a8c3bf387d0d6d88e669bd turns out not to have taken
care of all cases of the partially-evaluatable-PlaceHolderVar problem found
by Andreas Seltenreich's fuzz testing. I had set it up to check for risky
PHVs only in the event that we were making a star-schema-based exception to
the param_source_rels join ordering heuristic. However, it turns out that
the problem can occur even in joins that satisfy the param_source_rels
heuristic, in which case allow_star_schema_join() isn't consulted.
Refactor so that we check for risky PHVs whenever the proposed join has
any remaining parameterization.
Back-patch to 9.2, like the previous patch (except for the regression test
case, which only works back to 9.3 because it uses LATERAL).
Note that this discovery implies that problems of this sort could've
occurred in 9.2 and up even before the star-schema patch; though I've not
tried to prove that experimentally.
|
|
Further testing revealed that commit f69b4b9495269cc4 was still a few
bricks shy of a load: minor tweaking of the previous test cases resulted
in the same wrong-outer-join-order problem coming back. After study
I concluded that my previous changes in make_outerjoininfo() were just
accidentally masking the problem, and should be reverted in favor of
forcing syntactic join order whenever an upper outer join's predicate
doesn't mention a lower outer join's LHS. This still allows the
chained-outer-joins style that is the normally optimizable case.
I also tightened things up some more in join_is_legal(). It seems to me
on review that what's really happening in the exception case where we
ignore a mismatched special join is that we're allowing the proposed join
to associate into the RHS of the outer join we're comparing it to. As
such, we should *always* insist that the proposed join be a left join,
which eliminates a bunch of rather dubious argumentation. The case where
we weren't enforcing that was the one that was already known buggy anyway
(it had a violatable Assert before the aforesaid commit) so it hardly
deserves a lot of deference.
Back-patch to all active branches, like the previous patch. The added
regression test case failed in all branches back to 9.1, and I think it's
only an unrelated change in costing calculations that kept 9.0 from
choosing a broken plan.
|
|
Per the discussion in optimizer/README, it's unsafe to reassociate anything
into or out of the RHS of a SEMI or ANTI join. An example from Piotr
Stefaniak showed that join_is_legal() wasn't sufficiently enforcing this
rule, so lock it down a little harder.
I couldn't find a reasonably simple example of the optimizer trying to
do this, so no new regression test. (Piotr's example involved the random
search in GEQO accidentally trying an invalid case and triggering a sanity
check way downstream in clause selectivity estimation, which did not seem
like a sequence of events that would be useful to memorialize in a
regression test as-is.)
Back-patch to all active branches.
|
|
Per discussion, it really ought to do this. The original choice to
exclude shell types was probably made in the dark ages before we made
it harder to accidentally create shell types; but that was in 7.3.
Also, cause the standard regression tests to leave a shell type behind,
for convenience in testing the case in pg_dump and pg_upgrade.
Back-patch to all supported branches.
|
|
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.
|
|
In commit b514a7460d9127ddda6598307272c701cbb133b7, I changed the planner
so that it would allow nestloop paths to remain partially parameterized,
ie the inner relation might need parameters from both the current outer
relation and some upper-level outer relation. That's fine so long as we're
talking about distinct parameters; but the patch also allowed creation of
nestloop paths for cases where the inner relation's parameter was a
PlaceHolderVar whose eval_at set included the current outer relation and
some upper-level one. That does *not* work.
In principle we could allow such a PlaceHolderVar to be evaluated at the
lower join node using values passed down from the upper relation along with
values from the join's own outer relation. However, nodeNestloop.c only
supports simple Vars not arbitrary expressions as nestloop parameters.
createplan.c is also a few bricks shy of being able to handle such cases;
it misplaces the PlaceHolderVar parameters in the plan tree, which is why
the visible symptoms of this bug are "plan should not reference subplan's
variable" and "failed to assign all NestLoopParams to plan nodes" planner
errors.
Adding the necessary complexity to make this work doesn't seem like it
would be repaid in significantly better plans, because in cases where such
a PHV exists, there is probably a corresponding join order constraint that
would allow a good plan to be found without using the star-schema exception.
Furthermore, adding complexity to nodeNestloop.c would create a run-time
penalty even for plans where this whole consideration is irrelevant.
So let's just reject such paths instead.
Per fuzz testing by Andreas Seltenreich; the added regression test is based
on his example query. Back-patch to 9.2, like the previous patch.
|
|
It must be possible to multiply wal_buffers by XLOG_BLCKSZ without
overflowing int, or calculations in StartupXLOG will go badly wrong
and crash the server. Avoid that by imposing a maximum value on
wal_buffers. This will be just under 2GB, assuming the usual value
for XLOG_BLCKSZ.
Josh Berkus, per an analysis by Andrew Gierth.
|
|
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.
|
|
An outer join clause that didn't actually reference the RHS (perhaps only
after constant-folding) could confuse the join order enforcement logic,
leading to wrong query results. Also, nested occurrences of such things
could trigger an Assertion that on reflection seems incorrect.
Per fuzz testing by Andreas Seltenreich. The practical use of such cases
seems thin enough that it's not too surprising we've not heard field
reports about it.
This has been broken for a long time, so back-patch to all active branches.
|
|
Although I think on all modern machines floating division by zero
results in Infinity not SIGFPE, we still don't want infinities
running around in the planner's costing estimates; too much risk
of that leading to insane behavior.
grouping_planner() failed to consider the possibility that final_rel
might be known dummy and hence have zero rowcount. (I wonder if it
would be better to set a rows estimate of 1 for dummy relations?
But at least in the back branches, changing this convention seems
like a bad idea, so I'll leave that for another day.)
Make certain that get_variable_numdistinct() produces a nonzero result.
The case that can be shown to be broken is with stadistinct < 0.0 and
small ntuples; we did not prevent the result from rounding to zero.
For good luck I applied clamp_row_est() to all the nonconstant return
values.
In ExecChooseHashTableSize(), Assert that we compute positive nbuckets
and nbatch. I know of no reason to think this isn't the case, but it
seems like a good safety check.
Per reports from Piotr Stefaniak. Back-patch to all active branches.
|
|
Don't print a WARNING if we get ESRCH from a kill() that's attempting
to cancel an autovacuum worker. It's possible (and has been seen in the
buildfarm) that the worker is already gone by the time we are able to
execute the kill, in which case the failure is harmless. About the only
plausible reason for reporting such cases would be to help debug corrupted
lock table contents, but this is hardly likely to be the most important
symptom if that happens. Moreover issuing a WARNING might scare users
more than is warranted.
Also, since sending a signal to an autovacuum worker is now entirely a
routine thing, and the worker will log the query cancel on its end anyway,
reduce the message saying we're doing that from LOG to DEBUG1 level.
Very minor cosmetic cleanup as well.
Since the main practical reason for doing this is to avoid unnecessary
buildfarm failures, back-patch to all active branches.
|
|
While postgres' use of SSL renegotiation is a good idea in theory, it
turned out to not work well in practice. The specification and openssl's
implementation of it have lead to several security issues. Postgres' use
of renegotiation also had its share of bugs.
Additionally OpenSSL has a bunch of bugs around renegotiation, reported
and open for years, that regularly lead to connections breaking with
obscure error messages. We tried increasingly complex workarounds to get
around these bugs, but we didn't find anything complete.
Since these connection breakages often lead to hard to debug problems,
e.g. spuriously failing base backups and significant latency spikes when
synchronous replication is used, we have decided to change the default
setting for ssl renegotiation to 0 (disabled) in the released
backbranches and remove it entirely in 9.5 and master..
Author: Michael Paquier, with changes by me
Discussion: 20150624144148.GQ4797@alap3.anarazel.de
Backpatch: 9.0-9.4; 9.5 and master get a different patch
|
|
join_clause_is_movable_into() is approximate, in the sense that it might
sometimes return "false" when actually it would be valid to push the given
join clause down to the specified level. This is okay ... but there was
an Assert in get_joinrel_parampathinfo() that's only safe if the answers
are always exact. Comment out the Assert, and add a bunch of commentary
to clarify what's going on.
Per fuzz testing by Andreas Seltenreich. The added regression test is
a pretty silly query, but it's based on his crasher example.
Back-patch to 9.2 where the faulty logic was introduced.
|
|
It does currently, and I don't see us changing that any time soon, but we
don't make that assumption anywhere else.
Per Tom Lane's suggestion. Backpatch to 9.2, like the previous patch that
added this assumption.
|
|
In GIN, an all-zeros page would be leaked forever, and never reused. Just
add them to the FSM in vacuum, and they will be reinitialized when grabbed
from the FSM. On master and 9.5, attempting to access the page's opaque
struct also caused an assertion failure, although that was otherwise
harmless.
Reported by Jeff Janes. Backpatch to all supported versions.
|
|
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.
|
|
The planner generally expects that the estimated rowcount of any relation
is at least one row, *unless* it has been proven empty by constraint
exclusion or similar mechanisms, which is marked by installing a dummy path
as the rel's cheapest path (cf. IS_DUMMY_REL). When I split up
allpaths.c's processing of base rels into separate set_base_rel_sizes and
set_base_rel_pathlists steps, the intention was that dummy rels would get
marked as such during the "set size" step; this is what justifies an Assert
in indxpath.c's get_loop_count that other relations should either be dummy
or have positive rowcount. Unfortunately I didn't get that quite right
for append relations: if all the child rels have been proven empty then
set_append_rel_size would come up with a rowcount of zero, which is
correct, but it didn't then do set_dummy_rel_pathlist. (We would have
ended up with the right state after set_append_rel_pathlist, but that's
too late, if we generate indexpaths for some other rel first.)
In addition to fixing the actual bug, I installed an Assert enforcing this
convention in set_rel_size; that then allows simplification of a couple
of now-redundant tests for zero rowcount in set_append_rel_size.
Also, to cover the possibility that third-party FDWs have been careless
about not returning a zero rowcount estimate, apply clamp_row_est to
whatever an FDW comes up with as the rows estimate.
Per report from Andreas Seltenreich. Back-patch to 9.2. Earlier branches
did not have the separation between set_base_rel_sizes and
set_base_rel_pathlists steps, so there was no intermediate state where an
appendrel would have had inconsistent rowcount and pathlist. It's possible
that adding the Assert to set_rel_size would be a good idea in older
branches too; but since they're not under development any more, it's likely
not worth the trouble.
|
|
This was broken by commit 0e7e355f27302b62af3e1add93853ccd45678443 and
friends, which ignored the fact that gzopen() will treat "-1" in the
mode argument as an invalid character, which it ignores, and a flag for
compression level 1. Now, when this value is encountered no compression
level flag is passed to gzopen, leaving it to use the zlib default.
Also, enforce the documented allowed range for pg_dump's -Z option,
namely 0 .. 9, and remove some consequently dead code from
pg_backup_tar.c.
Problem reported by Marc Mamin.
Backpatch to 9.1, like the patch that introduced the bug.
|
|
If there were no subtransactions (or multixacts) active, we would calculate
the oldestxid == next xid. That's correct, but if next XID happens to be
on the next pg_subtrans (pg_multixact) page, the page does not exist yet,
and SimpleLruTruncate will produce an "apparent wraparound" warning. The
warning is harmless in this case, but looks very alarming to users.
Backpatch to all supported versions. Patch and analysis by Thomas Munro.
|
|
As reported by Bill Parker, PL/Tcl did not validate some malloc() calls
against NULL return. Fix by using palloc() in a new long-lived memory
context instead. This allows us to simplify error handling too, by
simply deleting the memory context instead of doing retail frees.
There's still a lot that could be done to improve PL/Tcl's memory
handling ...
This is pretty ancient, so backpatch all the way back.
Author: Michael Paquier and Álvaro Herrera
Discussion: https://www.postgresql.org/message-id/CAFrbyQwyLDYXfBOhPfoBGqnvuZO_Y90YgqFM11T2jvnxjLFmqw@mail.gmail.com
|
|
In the previous coding, timeout would be noticed and reported only when
poll() or socket() returned zero (or the equivalent behavior on Windows).
Ordinarily that should work well enough, but it seems conceivable that we
could get into a state where poll() always returns a nonzero value --- for
example, if it is noticing a condition on one of the file descriptors that
we do not think is reason to exit the loop. If that happened, we'd be in a
busy-wait loop that would fail to terminate even when the timeout expires.
We can make this more robust at essentially no cost, by deciding to exit
of our own accord if we compute a zero or negative time-remaining-to-wait.
Previously the code noted this but just clamped the time-remaining to zero,
expecting that we'd detect timeout on the next loop iteration.
Back-patch to 9.2. While 9.1 had a version of WaitLatchOrSocket, it was
primitive compared to later versions, and did not guarantee reliable
detection of timeouts anyway. (Essentially, this is a refinement of
commit 3e7fdcffd6f77187, which was back-patched only as far as 9.2.)
|
|
xlc provides "long long" unconditionally at C99-compatible language
levels, and this option provokes a warning. The warning interferes with
"configure" tests that fail in response to any warning. Notably, before
commit 85a2a8903f7e9151793308d0638621003aded5ae, it interfered with the
test for -qnoansialias. Back-patch to 9.0 (all supported versions).
|
|
It's standard for quicksort implementations, after having partitioned the
input into two subgroups, to recurse to process the smaller partition and
then handle the larger partition by iterating. This method guarantees
that no more than log2(N) levels of recursion can be needed. However,
Bentley and McIlroy argued that checking to see which partition is smaller
isn't worth the cycles, and so their code doesn't do that but just always
recurses on the left partition. In most cases that's fine; but with
worst-case input we might need O(N) levels of recursion, and that means
that qsort could be driven to stack overflow. Such an overflow seems to
be the only explanation for today's report from Yiqing Jin of a SIGSEGV
in med3_tuple while creating an index of a couple billion entries with a
very large maintenance_work_mem setting. Therefore, let's spend the few
additional cycles and lines of code needed to choose the smaller partition
for recursion.
Also, fix up the qsort code so that it properly uses size_t not int for
some intermediate values representing numbers of items. This would only
be a live risk when sorting more than INT_MAX bytes (in qsort/qsort_arg)
or tuples (in qsort_tuple), which I believe would never happen with any
caller in the current core code --- but perhaps it could happen with
call sites in third-party modules? In any case, this is trouble waiting
to happen, and the corrected code is probably if anything shorter and
faster than before, since it removes sign-extension steps that had to
happen when converting between int and size_t.
In passing, move a couple of CHECK_FOR_INTERRUPTS() calls so that it's
not necessary to preserve the value of "r" across them, and prettify
the output of gen_qsort_tuple.pl a little.
Back-patch to all supported branches. The odds of hitting this issue
are probably higher in 9.4 and up than before, due to the new ability
to allocate sort workspaces exceeding 1GB, but there's no good reason
to believe that it's impossible to crash older branches this way.
|
|
This allows PostgreSQL modules and their dependencies to have undefined
symbols, resolved at runtime. Perl module shared objects rely on that
in Perl 5.8.0 and later. This fixes the crash when PL/PerlU loads such
modules, as the hstore_plperl test suite does. Module authors can link
using -Wl,-G to permit undefined symbols; by default, linking will fail
as it has. Back-patch to 9.0 (all supported versions).
|
|
Per Coverity (not that any of these are so non-obvious that they should not
have been caught before commit). The extent of leakage is probably minor
to unnoticeable, but a leak is a leak. Back-patch as necessary.
Michael Paquier
|
|
Ordinarily, a failure (unexpected exit status) of the startup subprocess
should be considered fatal, so the postmaster should just close up shop
and quit. However, if we sent the startup process a SIGQUIT or SIGKILL
signal, the failure is hardly "unexpected", and we should attempt restart;
this is necessary for recovery from ordinary backend crashes in hot-standby
scenarios. I attempted to implement the latter rule with a two-line patch
in commit 442231d7f71764b8c628044e7ce2225f9aa43b67, but it now emerges that
that patch was a few bricks shy of a load: it failed to distinguish the
case of a signaled startup process from the case where the new startup
process crashes before reaching database consistency. That resulted in
infinitely respawning a new startup process only to have it crash again.
To handle this properly, we really must track whether we have sent the
*current* startup process a kill signal. Rather than add yet another
ad-hoc boolean to the postmaster's state, I chose to unify this with the
existing RecoveryError flag into an enum tracking the startup process's
state. That seems more consistent with the postmaster's general state
machine design.
Back-patch to 9.0, like the previous patch.
|
|
The psql crash happened when no current connection existed. (The second
new check is optional given today's undocumented NULL argument handling
in PQhost() etc.) Back-patch to 9.0 (all supported versions).
|
|
If an allocation fails in the main message handling loop, pqParseInput3
or pqParseInput2, it should not be treated as "not enough data available
yet". Otherwise libpq will wait indefinitely for more data to arrive from
the server, and gets stuck forever.
This isn't a complete fix - getParamDescriptions and getCopyStart still
have the same issue, but it's a step in the right direction.
Michael Paquier and me. Backpatch to all supported versions.
|
|
Build.bat and vcregress.bat got similar treatment years ago. I'm not sure
why install.bat wasn't treated at the same time, but it seems like a good
idea anyway.
The immediate problem with the old install.bat was that it had quoting
issues, and wouldn't work if the target directory's name contained spaces.
This fixes that problem.
I committed this to master yesterday, this is a backpatch of the same for
all supported versions.
|
|
Expose PG_VERSION_NUM (e.g., "90600") as a Make variable; but for
consistency with the other Make variables holding similar info,
call the variable just VERSION_NUM not PG_VERSION_NUM.
There was some discussion of making this value available as a pg_config
value as well. However, that would entail substantially more work than
this two-line patch. Given that there was not exactly universal consensus
that we need this at all, let's just do a minimal amount of work for now.
Back-patch of commit a5d489ccb7e613c7ca3be6141092b8c1d2c13fa7, so that this
variable is actually useful for its intended purpose sometime before 2020.
Michael Paquier, reviewed by Pavel Stehule
|
|
|
|
|
|
VACUUM FREEZE generated false cancelations of standby queries on an
otherwise idle master. Caused by an off-by-one error on cutoff_xid
which goes back to original commit.
Backpatch to all versions 9.0+
Analysis and report by Marco Nenciarini
Bug fix by Simon Riggs
|
|
Commit f3b5565dd4e59576be4c772da364704863e6a835 was a couple of bricks shy
of a load; specifically, it missed putting pg_trigger_tgrelid_tgname_index
into the relcache init file, because that index is not used by any
syscache. However, we have historically nailed that index into cache for
performance reasons. The upshot was that load_relcache_init_file always
decided that the init file was busted and silently ignored it, resulting
in a significant hit to backend startup speed.
To fix, reinstantiate RelationIdIsInInitFile() as a wrapper around
RelationSupportsSysCache(), which can know about additional relations
that should be in the init file despite being unknown to syscache.c.
Also install some guards against future mistakes of this type: make
write_relcache_init_file Assert that all nailed relations get written to
the init file, and make load_relcache_init_file emit a WARNING if it takes
the "wrong number of nailed relations" exit path. Now that we remove the
init files during postmaster startup, that case should never occur in the
field, even if we are starting a minor-version update that added or removed
rels from the nailed set. So the warning shouldn't ever be seen by end
users, but it will show up in the regression tests if somebody breaks this
logic.
Back-patch to all supported branches, like the previous commit.
|
|
Commit c03ad5602f529787968fa3201b35c119bbc6d782 introduced a planner
performance regression for UPDATE/DELETE on large inheritance sets.
It required copying the append_rel_list (which is of size proportional to
the number of inherited tables) once for each inherited table, thus
resulting in O(N^2) time and memory consumption. While it's difficult to
avoid that in general, the extra work only has to be done for
append_rel_list entries that actually reference subquery RTEs, which
inheritance-set entries will not. So we can buy back essentially all of
the loss in cases without subqueries in FROM; and even for those, the added
work is mainly proportional to the number of UNION ALL subqueries.
Back-patch to 9.2, like the previous commit.
Tom Lane and Dean Rasheed, per a complaint from Thomas Munro.
|
|
This supplements the GNU libc bug #6530 workarounds introduced in commit
54cd4f04576833abc394e131288bf3dd7dcf4806. On affected systems, a
tar-format pg_basebackup failed when some filename beneath the data
directory was not valid character data in the postmaster/walsender
locale. Back-patch to 9.1, where pg_basebackup was introduced. Extant,
bug-prone conversion specifications receive only ASCII bytes or involve
low-importance messages.
|
|
|
|
This avoids the problem that it might go to sleep for an unreasonable
amount of time in unusual conditions like the server clock moving
backwards an unreasonable amount of time.
(Simply moving the server clock forward again doesn't solve the problem
unless you wake up the autovacuum launcher manually, say by sending it
SIGHUP).
Per trouble report from Prakash Itnal in
https://www.postgresql.org/message-id/CAHC5u79-UqbapAABH2t4Rh2eYdyge0Zid-X=Xz-ZWZCBK42S0Q@mail.gmail.com
Analyzed independently by Haribabu Kommi and Tom Lane.
|
|
Patch by Michael Paquier
|
|
Patch by Michael Paquier
|
|
Patch by Michael Paquier
Conflicts:
src/interfaces/ecpg/preproc/variable.c
|
|
Patch by Michael Paquier
|
|
We already tried to improve this once, but the "improved" text was rather
off-target if you had provided a USING clause. Also, it seems helpful
to provide the exact text of a suggested USING clause, so users can just
copy-and-paste it when needed. Per complaint from Keith Rarick and a
suggestion from Merlin Moncure.
Back-patch to 9.2 where the current wording was adopted.
|
|
|
|
We don't know why a few Windows users have seen this fail, but the
taciturnity of the error message certainly isn't helping debug it.
Let's at least find out which LC category isn't working.
|
|
HotStandbyActiveInReplay, introduced in 061b079f, only allowed WAL
replay to happen in the startup process, missing the single user case.
This buglet is fairly harmless as it only causes problems when single
user mode in an assertion enabled build is used to replay a btree vacuum
record.
Backpatch to 9.2. 061b079f was backpatched further, but the assertion
was not.
|
|
When we invalidate the relcache entry for a system catalog or index, we
must also delete the relcache "init file" if the init file contains a copy
of that rel's entry. The old way of doing this relied on a specially
maintained list of the OIDs of relations present in the init file: we made
the list either when reading the file in, or when writing the file out.
The problem is that when writing the file out, we included only rels
present in our local relcache, which might have already suffered some
deletions due to relcache inval events. In such cases we correctly decided
not to overwrite the real init file with incomplete data --- but we still
used the incomplete initFileRelationIds list for the rest of the current
session. This could result in wrong decisions about whether the session's
own actions require deletion of the init file, potentially allowing an init
file created by some other concurrent session to be left around even though
it's been made stale.
Since we don't support changing the schema of a system catalog at runtime,
the only likely scenario in which this would cause a problem in the field
involves a "vacuum full" on a catalog concurrently with other activity, and
even then it's far from easy to provoke. Remarkably, this has been broken
since 2002 (in commit 786340441706ac1957a031f11ad1c2e5b6e18314), but we had
never seen a reproducible test case until recently. If it did happen in
the field, the symptoms would probably involve unexpected "cache lookup
failed" errors to begin with, then "could not open file" failures after the
next checkpoint, as all accesses to the affected catalog stopped working.
Recovery would require manually removing the stale "pg_internal.init" file.
To fix, get rid of the initFileRelationIds list, and instead consult
syscache.c's list of relations used in catalog caches to decide whether a
relation is included in the init file. This should be a tad more efficient
anyway, since we're replacing linear search of a list with ~100 entries
with a binary search. It's a bit ugly that the init file contents are now
so directly tied to the catalog caches, but in practice that won't make
much difference.
Back-patch to all supported branches.
|
|
We should set MyProc->databaseId after acquiring the per-database lock,
not beforehand. The old way risked deadlock against processes trying to
copy or delete the target database, since they would first acquire the lock
and then wait for processes with matching databaseId to exit; that left a
window wherein an incoming process could set its databaseId and then block
on the lock, while the other process had the lock and waited in vain for
the incoming process to exit.
CountOtherDBBackends() would time out and fail after 5 seconds, so this
just resulted in an unexpected failure not a permanent lockup, but it's
still annoying when it happens. A real-world example of a use-case is that
short-duration connections to a template database should not cause CREATE
DATABASE to fail.
Doing it in the other order should be fine since the contract has always
been that processes searching the ProcArray for a database ID must hold the
relevant per-database lock while searching. Thus, this actually removes
the former race condition that required an assumption that storing to
MyProc->databaseId is atomic.
It's been like this for a long time, so back-patch to all active branches.
|
|
|
|
The argument that this is a sufficiently-expected case to be silently
ignored seems pretty thin. Andres had brought it up back when we were
still considering that most fsync failures should be hard errors, and it
probably would be legit not to fail hard for ETXTBSY --- but the same is
true for EROFS and other cases, which is why we gave up on hard failures.
ETXTBSY is surely not a normal case, so logging the failure seems fine
from here.
|
