Per post-commit review comments from Andres Freund, improve variable
names, comments, and in one place, slightly improve the code structure.
Masahiko Sawada
Use MAXALIGN size/alignment to guarantee that later uses of memory are
aligned correctly. E.g. epoll_event might need 8 byte alignment on some
platforms, but earlier allocations like WaitEventSet and WaitEvent might
not sized to guarantee that when purely using sizeof().
Found by myself while testing on an Sun Ultra 5 (Sparc IIi) with some
editorializing by Andres Freund.
In passing fix a couple typos in the area
Formerly, Unix builds of pg_dump/pg_restore would trap SIGINT and similar
signals and set a flag that was tested in various data-transfer loops.
This was prone to errors of omission (cf commit 3c8aa6654); and even if
the client-side response was prompt, we did nothing that would cause
long-running SQL commands (e.g. CREATE INDEX) to terminate early.
Also, the master process would effectively do nothing at all upon receipt
of SIGINT; the only reason it seemed to work was that in typical scenarios
the signal would also be delivered to the child processes. We should
support termination when a signal is delivered only to the master process,
though.
Windows builds had no console interrupt handler, so they would just fall
over immediately at control-C, again leaving long-running SQL commands to
finish unmolested.
To fix, remove the flag-checking approach altogether. Instead, allow the
Unix signal handler to send a cancel request directly and then exit(1).
In the master process, also have it forward the signal to the children.
On Windows, add a console interrupt handler that behaves approximately
the same. The main difference is that a single execution of the Windows
handler can send all the cancel requests since all the info is available
in one process, whereas on Unix each process sends a cancel only for its
own database connection.
In passing, fix an old problem that DisconnectDatabase tends to send a
cancel request before exiting a parallel worker, even if nothing went
wrong. This is at least a waste of cycles, and could lead to unexpected
log messages, or maybe even data loss if it happened in pg_restore (though
in the current code the problem seems to affect only pg_dump). The cause
was that after a COPY step, pg_dump was leaving libpq in PGASYNC_BUSY
state, causing PQtransactionStatus() to report PQTRANS_ACTIVE. That's
normally harmless because the next PQexec() will silently clear the
PGASYNC_BUSY state; but in a parallel worker we might exit without any
additional SQL commands after a COPY step. So add an extra PQgetResult()
call after a COPY to allow libpq to return to PGASYNC_IDLE state.
This is a bug fix, IMO, so back-patch to 9.3 where parallel dump/restore
were introduced.
Thanks to Kyotaro Horiguchi for Windows testing and code suggestions.
Original-Patch: <7005.1464657274@sss.pgh.pa.us>
Discussion: <20160602.174941.256342236.horiguchi.kyotaro@lab.ntt.co.jp>
Commit 2ed5b87f96 introduced a bug in
mark/restore, in an attempt to optimize repeated restores to the
same page. This caused an assertion failure during a merge join
which fed directly from an index scan, although the impact would
not be limited to that case. Revert the bad chunk of code from
that commit.
While investigating this bug it was discovered that a particular
"paranoia" set of the mark position field would not prevent bad
behavior; it would just make it harder to diagnose. Change that
into an assertion, which will draw attention to any future problem
in that area more directly.
Backpatch to 9.5, where the bug was introduced.
Bug #14169 reported by Shinta Koyanagi.
Preliminary analysis by Tom Lane identified which commit caused
the bug.
Parallel dump did a totally pointless query to find out the name of each
table to be dumped, which it already knows. Parallel restore runs issued
lots of redundant SET commands because _doSetFixedOutputState() was invoked
once per TOC item rather than just once at connection start. While the
extra queries are insignificant if you're dumping or restoring large
tables, it still seems worth getting rid of them.
Also, give the responsibility for selecting the right client_encoding for
a parallel dump worker to setup_connection() where it naturally belongs,
instead of having ad-hoc code for that in CloneArchive(). And fix some
minor bugs like use of strdup() where pg_strdup() would be safer.
Back-patch to 9.3, mostly to keep the branches in sync in an area that
we're still finding bugs in.
Discussion: <5086.1464793073@sss.pgh.pa.us>
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 187492b6c2, 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>
This bug appears to have been introduced late in the development of
48354581a4 ("Allow Pin/UnpinBuffer to operate in a lockfree
manner.").
Found while debugging a bug which turned out to be independent of the
commit mentioned above.
Backpatch: -
BRIN was relying on the ability to remove a tuple from an index page,
then putting another tuple in the same line pointer. But PageAddItem
refuses to add a tuple beyond the first free item past the last used
item, and in particular, it rejects an attempt to add an item to an
empty page anywhere other than the first line pointer. PageAddItem
issues a WARNING and indicates to the caller that it failed, which in
turn causes the BRIN calling code to issue a PANIC, so the whole
sequence looks like this:
WARNING: specified item offset is too large
PANIC: failed to add BRIN tuple
To fix, create a new function PageAddItemExtended which is like
PageAddItem except that the two boolean arguments become a flags bitmap;
the "overwrite" and "is_heap" boolean flags in PageAddItem become
PAI_OVERWITE and PAI_IS_HEAP flags in the new function, and a new flag
PAI_ALLOW_FAR_OFFSET enables the behavior required by BRIN.
PageAddItem() retains its original signature, for compatibility with
third-party modules (other callers in core code are not modified,
either).
Also, in the belt-and-suspenders spirit, I added a new sanity check in
brinGetTupleForHeapBlock to raise an error if an TID found in the revmap
is not marked as live by the page header. This causes it to react with
"ERROR: corrupted BRIN index" to the bug at hand, rather than a hard
crash.
Backpatch to 9.5.
Bug reported by Andreas Seltenreich as detected by his handy sqlsmith
fuzzer.
Discussion: https://www.postgresql.org/message-id/87mvni77jh.fsf@elite.ansel.ydns.eu
Parallel restore from directory format failed to respond to control-C
in a timely manner, because there were no checkAborting() calls in the
code path that reads data from a file and sends it to the backend.
If any worker was in the midst of restoring data for a large table,
you'd just have to wait.
This fix doesn't do anything for the problem of aborting a long-running
server-side command, but at least it fixes things for data transfers.
Back-patch to 9.3 where parallel restore was introduced.
This was effectively dead code, since the places that tested it could not
be reached after we entered the on-exit-cleanup routine that would set it.
It seems to have been a leftover from a design in which error abort would
try to send fresh commands to the workers --- a design which could never
have worked reliably, of course. Since the flag is not cross-platform, it
complicates reasoning about the code's behavior, which we could do without.
Although this is effectively just cosmetic, back-patch anyway, because
there are some actual bugs in the vicinity of this behavior.
Discussion: <15583.1464462418@sss.pgh.pa.us>
The commentary in this file was in extremely sad shape. The author(s)
had clearly never heard of the project convention that a function header
comment should provide an API spec of some sort for that function. Much
of it was flat out wrong, too --- maybe it was accurate when written, but
if so it had not been updated to track subsequent code revisions. Rewrite
and rearrange to try to bring it up to speed, and annotate some of the
places where more work is needed. (I've refrained from actually fixing
anything of substance ... yet.)
Also, rename a couple of functions for more clarity as to what they do,
do some very minor code rearrangement, remove some pointless Asserts,
fix an incorrect Assert in readMessageFromPipe, and add a missing socket
close in one error exit from pgpipe(). The last would be a bug if we
tried to continue after pgpipe() failure, but since we don't, it's just
cosmetic at present.
Although this is only cosmetic, back-patch to 9.3 where parallel.c was
added. It's sufficiently invasive that it'll pose a hazard for future
back-patching if we don't.
Discussion: <25239.1464386067@sss.pgh.pa.us>
Since we start the worker threads with _beginthreadex(), we should use
_endthreadex() to terminate them. We got this right in the normal-exit
code path, but not so much during an error exit from a worker.
In addition, be sure to apply CloseHandle to the thread handle after
each thread exits.
It's not clear that these oversights cause any user-visible problems,
since the pg_dump run is about to terminate anyway. Still, it's clearly
better to follow Microsoft's API specifications than ignore them.
Also a few cosmetic cleanups in WaitForTerminatingWorkers(), including
being a bit less random about where to cast between uintptr_t and HANDLE,
and being sure to clear the worker identity field for each dead worker
(not that false matches should be possible later, but let's be careful).
Original observation and patch by Armin Schöffmann, cosmetic improvements
by Michael Paquier and me. (Armin's patch also included closing sockets
in ShutdownWorkersHard(), but that's been dealt with already in commit
df8d2d8c4.) Back-patch to 9.3 where parallel pg_dump was introduced.
Discussion: <zarafa.570306bd.3418.074bf1420d8f2ba2@root.aegaeon.de>
The IF EXISTS option was documented, and implemented in the grammar, but
it didn't actually work for lack of support in does_not_exist_skipping().
Per bug #14160.
Report and patch by Kouhei Sutou
Report: <20160527070433.19424.81712@wrigleys.postgresql.org>
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>
Getting a synchronized snapshot is not supported on a hot standby node,
and is by default taken when using -j with multiple sessions. Trying to
do so still failed, but with a server error that would also go in the
log. Instead, proprely detect this case and give a better error message.
As part of upper planner pathification (commit 3fc6e2d7f5) I redid
createplan.c's approach to the physical-tlist optimization, in which scan
nodes are allowed to return exactly the underlying table's columns so as
to save doing a projection step at runtime. The logic was intentionally
more aggressive than before about applying the optimization, which is
generally a good thing, but Andres Freund found a case in which it got
too aggressive. Namely, if any column is referenced more than once in
the parent plan node's sorting or grouping column list, we can't optimize
because then that column would need to have more than one ressortgroupref
label, and we only have space for one.
Add logic to detect this situation in use_physical_tlist(), and also add
some error checking in apply_pathtarget_labeling_to_tlist(), which this
example proves was being overly cavalier about whether what it was doing
made any sense.
The added test case exposes the problem only because we do not eliminate
duplicate grouping keys. That might be something to fix someday, but it
doesn't seem like appropriate post-beta work.
Report: <20160526021235.w4nq7k3gnheg7vit@alap3.anarazel.de>
For some reason the code to emit a warning and switch to uncompressed
output was placed down in the guts of pg_backup_archiver.c. This is
definitely too late in the case of parallel operation (and I rather
wonder if it wasn't too late for other purposes as well). Put it in
pg_dump.c's option-processing logic, which seems a much saner place.
Also, the default behavior with custom or directory output format was
to emit the warning telling you the output would be uncompressed. This
seems unhelpful, so silence that case.
Back-patch to 9.3 where parallel dump was introduced.
Kyotaro Horiguchi, adjusted a bit by me
Report: <20160526.185551.242041780.horiguchi.kyotaro@lab.ntt.co.jp>
The Windows coding of ShutdownWorkersHard() thought that setting termEvent
was sufficient to make workers exit after an error. But that only helps
if a worker is busy and passes through checkAborting(). An idle worker
will just sit, resulting in pg_dump failing to exit until the user gives up
and hits control-C. We should close the write end of the command pipe
so that idle workers will see socket EOF and exit, as the Unix coding was
already doing.
Back-patch to 9.3 where parallel pg_dump was introduced.
Kyotaro Horiguchi
Dating back to commit f10b63923, our grammar has allowed "USING" to
optionally appear before an opclass name in CREATE INDEX (and, lately,
some related places such as ON CONFLICT specifications). Nikolay Shaplov
noticed that this syntax existed but wasn't documented, and proposed
documenting it. But what seems like a better idea is to remove the
production, thereby making the code match the docs not vice versa.
This isn't our usual modus operandi for such cases, but there are a
couple of good reasons to proceed this way:
* So far as I can find, this syntax has never been documented anywhere.
It isn't relied on by any of our own code or test cases, and there seems
little reason to suppose that it's been used in the wild either.
* Documenting it would mean that there would be two separate uses of
USING in the CREATE INDEX syntax, the other being "USING access_method".
That can lead to nothing but confusion.
So, let's just remove it. On the off chance that somebody somewhere
is using it, this isn't something to back-patch, but we can fix it
in HEAD.
Discussion: <1593237.l7oKHRpxSe@nataraj-amd64>
Ever since we split the statistics collector's reports into per-database
files (commit 187492b6c2), backends have been seeing stale statistics
for shared catalogs. This is because the inquiry message only prompts the
collector to write the per-database file for the requesting backend's own
database. Stats for shared catalogs are in a separate file for "DB 0",
which didn't get updated.
In normal operation this was partially masked by the fact that the
autovacuum launcher would send an inquiry message at least once per
autovacuum_naptime that asked for "DB 0"; so the shared-catalog stats would
never be more than a minute out of date. However the problem becomes very
obvious with autovacuum disabled, as reported by Peter Eisentraut.
To fix, redefine the semantics of inquiry messages so that both the
specified DB and DB 0 will be dumped. (This might seem a bit inefficient,
but we have no good way to know whether a backend's transaction will look
at shared-catalog stats, so we have to read both groups of stats whenever
we request stats. Sending two inquiry messages would definitely not be
better.)
Back-patch to 9.3 where the bug was introduced.
Report: <56AD41AC.1030509@gmx.net>
In the original design for parallel dump, worker processes reported errors
by sending them up to the master process, which would print the messages.
This is unworkably fragile for a couple of reasons: it risks deadlock if a
worker sends an error at an unexpected time, and if the master has already
died for some reason, the user will never get to see the error at all.
Revert that idea and go back to just always printing messages to stderr.
This approach means that if all the workers fail for similar reasons (eg,
bad password or server shutdown), the user will see N copies of that
message, not only one as before. While that's slightly annoying, it's
certainly better than not seeing any message; not to mention that we
shouldn't assume that only the first failure is interesting.
An additional problem in the same area was that the master failed to
disable SIGPIPE (at least until much too late), which meant that sending a
command to an already-dead worker would cause the master to crash silently.
That was bad enough in itself but was made worse by the total reliance on
the master to print errors: even if the worker had reported an error, you
would probably not see it, depending on timing. Instead disable SIGPIPE
right after we've forked the workers, before attempting to send them
anything.
Additionally, the master relies on seeing socket EOF to realize that a
worker has exited prematurely --- but on Windows, there would be no EOF
since the socket is attached to the process that includes both the master
and worker threads, so it remains open. Make archive_close_connection()
close the worker end of the sockets so that this acts more like the Unix
case. It's not perfect, because if a worker thread exits without going
through exit_nicely() the closures won't happen; but that's not really
supposed to happen.
This has been wrong all along, so back-patch to 9.3 where parallel dump
was introduced.
Report: <2458.1450894615@sss.pgh.pa.us>
When pulling the list of roles to drop, exclude roles whose names
begin with "pg_" (as we do when we are dumping the roles out to
recreate them).
Also add regression tests to cover pg_dumpall -c and this specific
issue.
Noticed by Rushabh Lathia. Patch by me.
blbuildempty did not do even approximately the right thing: it tried
to add a metapage to the relation's regular data fork, which already
has one at that point. It should look like the ambuildempty methods
for all the standard index types, ie, initialize a metapage image in
some transient storage and then write it directly to the init fork.
To support that, refactor BloomInitMetapage into two functions.
In passing, fix BloomInitMetapage so it doesn't leave the rd_options
field of the index's relcache entry pointing at transient storage.
I'm not sure this had any visible consequence, since nothing much
else is likely to look at a bloom index's rd_options, but it's
certainly poor practice.
Per bug #14155 from Zhou Digoal.
Report: <20160524144146.22598.42558@wrigleys.postgresql.org>
All of the other tables used in the query in dumpTable(), which is
collecting column-level ACLs, are qualified, so we should be qualifying
the pg_init_privs, the related sub-select against pg_class and the
other queries added by the pg_dump catalog ACLs work.
Also, use ::regclass (or ::pg_catalog.regclass, where appropriate)
instead of using a poorly constructed query to get the OID for various
catalog tables.
Issues identified by Noah and Alvaro, patch by me.
Because vac_update_datfrozenxid() updates datfrozenxid and datminmxid
in-place, it's unsafe to assume that successive reads of those values will
give consistent results. Fetch each one just once to ensure sane behavior
in the minimum calculation. Noted while reviewing Alexander Korotkov's
patch in the same area.
Discussion: <8564.1464116473@sss.pgh.pa.us>
vac_truncate_clog() uses its own transaction ID as the comparison point in
a sanity check that no database's datfrozenxid has already wrapped around
"into the future". That was probably fine when written, but in a lazy
vacuum we won't have assigned an XID, so calling GetCurrentTransactionId()
causes an XID to be assigned when otherwise one would not be. Most of the
time that's not a big problem ... but if we are hard up against the
wraparound limit, consuming XIDs during antiwraparound vacuums is a very
bad thing.
Instead, use ReadNewTransactionId(), which not only avoids this problem
but is in itself a better comparison point to test whether wraparound
has already occurred.
Report and patch by Alexander Korotkov. Back-patch to all versions.
Report: <CAPpHfdspOkmiQsxh-UZw2chM6dRMwXAJGEmmbmqYR=yvM7-s6A@mail.gmail.com>
Commit 1aba62ec moved the range check of that option form guc.c into
bufmgr.c, but introduced a bug by changing a >= 0.0 to > 0.0, which made
the value 0 no longer accepted. Put it back.
Reported by Jeff Janes, diagnosed by Tom Lane
Oracle recommends using VARCHAR2 not VARCHAR, allegedly because they might
someday change VARCHAR to be spec-compliant about distinguishing null from
empty string. (I'm not holding my breath, though.) Our examples of PL/SQL
code were using VARCHAR, which while not wrong is missing the pedagogical
opportunity to talk about converting Oracle type names to Postgres. So
switch the examples to use VARCHAR2, and add some text about what to do
with common Oracle type names like VARCHAR2 and NUMBER. (There is probably
more to be said here, but those are the ones I'm sure about offhand.)
Per suggestion from rapg12@gmail.com.
Discussion: <20160521140046.22591.24672@wrigleys.postgresql.org>
Commit 65c5fcd353 broke this by removing a
header include directive that is conditionally required. Add that back
to nbtree.c, with annotation to keep pgrminclude from re-breaking it.
Peter Geoghegan
Report: <CAM3SWZTNjHFYW_UG8bu0BnogqQ2HfsTgkzXLueuUhfTcYbu5HA@mail.gmail.com>
Needed for cases in which INSERT ... ON CONFLICT appears inside a
recursive CTE item. Per bug #14153 from Thomas Alton.
Patch by Peter Geoghegan, slightly adjusted by me
Report: <20160521232802.22598.13537@wrigleys.postgresql.org>
If RAW_EXPRESSION_COVERAGE_TEST is defined, do a no-op tree walk over
every basic DML statement submitted to parse analysis. If we'd had this
in place earlier, bug #14153 would have been caught by buildfarm testing.
The difficulty is that raw_expression_tree_walker() is only used in
limited cases involving CTEs (particularly recursive ones), so it's
very easy for an oversight in it to not be noticed during testing of a
seemingly-unrelated feature.
The type of error we can expect to catch with this is complete omission
of a node type from raw_expression_tree_walker(), and perhaps also
recursion into a field that doesn't contain a node tree, though that
would be an unlikely mistake. It won't catch failure to add new fields
that need to be recursed into, unfortunately.
I'll go enable this on one or two of my own buildfarm animals once
bug #14153 is dealt with.
Discussion: <27861.1464040417@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>
Correct obsolete install instructions, as noted by Daniel Gustafsson.
Clarify the test code's prerequisites.
Discussion: <88E617F2-7721-4C4E-84F4-886A2041C1D0@yesql.se>
David Johnston pointed out that the original text here had been obsoleted
by SQL:2008, which allowed ORDER BY in subqueries. We could weaken the
text to describe ORDER-BY-in-subqueries as an optional SQL feature that's
possibly unportable; but then the exact same statements would apply to
the alternative it's being compared to (ORDER-BY-in-aggregate-calls).
So really that would be pretty useless; let's just take out the sentence
entirely. Instead, point out the hazard that any extra processing in the
upper query might cause the subquery output order to be destroyed.
Discussion: <CAKFQuwbAX=iO9QbpN7_jr+BnUWm9FYX8WbEPUvG0p+nZhp6TZg@mail.gmail.com>
This was overlooked in commit 473b93287, which introduced DROP ACCESS
METHOD. Although that command is restricted to superusers, we don't want
even superusers dropping the built-in methods; "DROP ACCESS METHOD btree"
in particular is unrecoverable from. Pin these objects in the same way
that other initdb-created objects are pinned.
I chose to bump catversion for this fix. That's not absolutely necessary
perhaps, but it will ensure that no 9.6 production systems are missing
the pin entries.
