I somehow had assumed that in the spinlock (in turn possibly using
semaphores) based fallback atomics implementation 32 bit writes could be
done without a lock. As far as the write goes that's correct, since
postgres supports only platforms with single-copy atomicity for aligned
32bit writes. But writing without holding the spinlock breaks
read-modify-write operations like pg_atomic_compare_exchange_u32(),
since they'll potentially "miss" a concurrent write, which can't happen
in actual hardware implementations.
In 9.6+ when using the fallback atomics implementation this could lead
to buffer header locks not being properly marked as released, and
potentially some related state corruption. I don't see a related danger
in 9.5 (earliest release with the API), because pg_atomic_write_u32()
wasn't used in a concurrent manner there.
The state variable of local buffers, before this change, were
manipulated using pg_atomic_write_u32(), to avoid unnecessary
synchronization overhead. As that'd not be the case anymore, introduce
and use pg_atomic_unlocked_write_u32(), which does not correctly
interact with RMW operations.
This bug only caused issues when postgres is compiled on platforms
without atomics support (i.e. no common new platform), or when compiled
with --disable-atomics, which explains why this wasn't noticed in
testing.
Reported-By: Tom Lane
Discussion: <14947.1475690465@sss.pgh.pa.us>
Backpatch: 9.5-, where the atomic operations API was introduced.
0xc19c is not a valid UTF-8 byte sequence. It doesn't do any harm, AFAICS,
but it's surely not intentional. No backpatching though, just to be sure.
In the passing, also add a file header comment to the file, like the
UCS_to_SJIS.pl script would produce. (The file was originally created with
UCS_to_SJIS.pl, but has been modified by hand since then. That's
questionable, but I'll leave fixing that for later.)
Kyotaro Horiguchi
Discussion: <20160907.155050.233844095.horiguchi.kyotaro@lab.ntt.co.jp>
Recent Perl and/or new Linux distributions are starting to remove "." from
the @INC list by default. That breaks pg_rewind and ssl test suites, which
use helper perl modules that reside in the same directory. To fix, add the
current source directory explicitly to prove's include dir.
The vcregress.pl script probably also needs something like this, but I
wasn't able to remove '.' from @INC on Windows to test this, and don't want
to try doing that blindly.
Discussion: <20160908204529.flg6nivjuwp5vaoy@alap3.anarazel.de>
getBlobs' queries for pre-9.0 servers were broken in two ways:
the 7.x/8.x query uses DISTINCT so it can't have unspecified-type
NULLs in the target list, and both that query and the 7.0 one
failed to provide the correct output column labels, so that the
subsequent code to extract data from the PGresult would fail.
Back-patch to 9.6 where the breakage was introduced (by commit 23f34fa4b).
Amit Langote and Tom Lane
Discussion: <0a3e7a0e-37bd-8427-29bd-958135862f0a@lab.ntt.co.jp>
They are supposed to be mutually exclusive, but there was no check for
that.
Michael Banck
Discussion: <20161007103414.GD12247@nighthawk.caipicrew.dd-dns.de>
Leaving the error in the error queue used to be harmless, because the
X509_STORE_load_locations() call used to be the last step in
initialize_SSL(), and we would clear the queue before the next
SSL_connect() call. But previous commit moved things around. The symptom
was that if a CRL file was not found, and one of the subsequent
initialization steps, like loading the client certificate or private key,
failed, we would incorrectly print the "no such file" error message from
the earlier X509_STORE_load_locations() call as the reason.
Backpatch to all supported versions, like the previous patch.
There were several issues with the old coding:
1. There was a race condition, if two threads opened a connection at the
same time. We used a mutex around SSL_CTX_* calls, but that was not
enough, e.g. if one thread SSL_CTX_load_verify_locations() with one
path, and another thread set it with a different path, before the first
thread got to establish the connection.
2. Opening two different connections, with different sslrootcert settings,
seemed to fail outright with "SSL error: block type is not 01". Not sure
why.
3. We created the SSL object, before calling SSL_CTX_load_verify_locations
and SSL_CTX_use_certificate_chain_file on the SSL context. That was
wrong, because the options set on the SSL context are propagated to the
SSL object, when the SSL object is created. If they are set after the
SSL object has already been created, they won't take effect until the
next connection. (This is bug #14329)
At least some of these could've been fixed while still using a shared
context, but it would've been more complicated and error-prone. To keep
things simple, let's just use a separate SSL context for each connection,
and accept the overhead.
Backpatch to all supported versions.
Report, analysis and test case by Kacper Zuk.
Discussion: <20160920101051.1355.79453@wrigleys.postgresql.org>
If you point pg_rewind to a server that is using synchronous replication,
with "pg_rewind --source-server=...", and the replication is not working
for some reason, pg_rewind will get stuck because it creates a temporary
table, which needs to be replicated. You could call broken replication a
pilot error, but pg_rewind is often used in special circumstances, when
there are changes to the replication setup.
We don't do any "real" updates, and we don't care about fsyncing or
replicating the operations on the temporary tables, so fix that by
setting synchronous_commit off.
Michael Banck, Michael Paquier. Backpatch to 9.5, where pg_rewind was
introduced.
Discussion: <20161005143938.GA12247@nighthawk.caipicrew.dd-dns.de>
LogicalTapeRewind() should not allocate large read buffer, if the tape
is completely empty. The calling code relies on that, for its
calculation of how much memory to allocate for the read buffers. That
lead to massive overallocation of memory, if maxTapes was high, but
only a few tapes were actually used.
Reported by Tomas Vondra
Discussion: <7303da46-daf7-9c68-3cc1-9f83235cf37e@2ndquadrant.com>
We don't need this anymore, and it prevents build-time error checking
that's usually good to have, so remove it. Undoes one change of commit
cac765820.
Unfortunately, it's much harder to get a similar effect on other common
platforms, because we don't want the linker to throw errors for symbols
that will be resolved in the core backend. Only macOS and AIX expect the
core backend executable to be available while linking loadable modules,
so only these platforms can usefully throw errors for unresolved symbols
at link time.
Discussion: <2652.1475512158@sss.pgh.pa.us>
Windows apparently has a constant named WAIT_TIMEOUT, and some of these
other names are pretty generic, too. Insert "PG_" at the front of each
name in order to disambiguate.
Michael Paquier
Just turning the crank on the project started in commit d51924be8.
These cases turn out to be exact subsets of the boilerplate needed
for hstore_plpython.
Discussion: <2652.1475512158@sss.pgh.pa.us>
WaitLatch, WaitLatchOrSocket, and WaitEventSetWait now taken an
additional wait_event_info parameter; legal values are defined in
pgstat.h. This makes it possible to uniquely identify every point in
the core code where we are waiting for a latch; extensions can pass
WAIT_EXTENSION.
Because latches were the major wait primitive not previously covered
by this patch, it is now possible to see information in
pg_stat_activity on a large number of important wait events not
previously addressed, such as ClientRead, ClientWrite, and SyncRep.
Unfortunately, many of the wait events added by this patch will fail
to appear in pg_stat_activity because they're only used in background
processes which don't currently appear in pg_stat_activity. We should
fix this either by creating a separate view for such information, or
else by deciding to include them in pg_stat_activity after all.
Michael Paquier and Robert Haas, reviewed by Alexander Korotkov and
Thomas Munro.
Before initializing iteration over a subtransaction's changes, the last
few changes were not spilled to disk. That's correct if the transaction
didn't spill to disk, but otherwise... This bug can lead to missed or
misorderd subtransaction contents when they were spilled to disk.
Move spilling of the remaining in-memory changes to
ReorderBufferIterTXNInit(), where it can easily be applied to the top
transaction and, if present, subtransactions.
Since this code had too many bugs already, noticeably increase test
coverage.
Fixes: #14319
Reported-By: Huan Ruan
Discussion: <20160909012610.20024.58169@wrigleys.postgresql.org>
Backport: 9,4-, where logical decoding was added
Previously, on most platforms, we allowed hstore_plpython's references
to hstore and plpython to be unresolved symbols at link time, trusting
the dynamic linker to resolve them when the module is loaded. This
has a number of problems, the worst being that the dynamic linker
does not know where the references come from and can do nothing but
fail if those other modules haven't been loaded. We've more or less
gotten away with that for the limited use-case of datatype transform
modules, but even there, it requires some awkward hacks, most recently
commit 83c249200.
Instead, let's not treat these references as linker-resolvable at all,
but use function pointers that are manually filled in by the module's
_PG_init function. There are few enough contact points that this
doesn't seem unmaintainable, at least for these use-cases. (Note that
the same technique wouldn't work at all for decoupling from libpython
itself, but fortunately that's just a standard shared library and can
be linked to normally.)
This is an initial patch that just converts hstore_plpython. If the
buildfarm doesn't find any fatal problems, I'll work on the other
transform modules soon.
Tom Lane, per an idea of Andres Freund's.
Discussion: <2652.1475512158@sss.pgh.pa.us>
Commit 88e982302 invented GUC_UNIT_XSEGS for min_wal_size and max_wal_size,
but neglected to make it display sensibly in pg_settings.unit (by adding a
case to the switch in GetConfigOptionByNum). Fix that, and adjust said
switch to throw a run-time error the next time somebody forgets.
In passing, avoid using a static buffer for the output string --- the rest
of this function pstrdup's from a local buffer, and I see no very good
reason why the units code should do it differently and less safely.
Per report from Otar Shavadze. Back-patch to 9.5 where the new unit type
was added.
Report: <CAG-jOyA=iNFhN+yB4vfvqh688B7Tr5SArbYcFUAjZi=0Exp-Lg@mail.gmail.com>
Attempting to COPY a subset of columns from a table with RLS enabled
would fail due to an invalid query being constructed (using a single
ColumnRef with the list of fields to exact in 'fields', but that's for
the different levels of an indirection for a single column, not for
specifying multiple columns).
Correct by building a ColumnRef and then RestTarget for each column
being requested and then adding those to the targetList for the select
query. Include regression tests to hopefully catch if this is broken
again in the future.
Patch-By: Adam Brightwell
Reviewed-By: Michael Paquier
pg_upgrade checks whether all the shared libraries used in the old cluster
are also available in the new one by issuing LOAD for each library name.
Previously, it cared not what order it did the LOADs in. Ideally it
should not have to care, but currently the transform modules in contrib
fail unless both the language and datatype modules they depend on are
loaded first. A backend-side solution for that looks possible but
probably not back-patchable, so as a stopgap measure, let's do the LOAD
tests in order by library name length. That should fix the problem for
reasonably-named transform modules, eg "hstore_plpython" will be loaded
after both "hstore" and "plpython". (Yeah, it's a hack.)
In a larger sense, having a predictable order of these probes is a good
thing, since it will make upgrades predictably work or not work in the
face of inter-library dependencies. Also, this patch replaces O(N^2)
de-duplication logic with O(N log N) logic, which could matter in
installations with very many databases. So I don't foresee reverting this
even after we have a proper fix for the library-dependency problem.
In passing, improve a couple of SQL queries used here.
Per complaint from Andrew Dunstan that pg_upgrade'ing the transform contrib
modules failed. Back-patch to 9.5 where transform modules were introduced.
Discussion: <f7ac29f3-515c-2a44-21c5-ec925053265f@dunslane.net>
Don't pre-read tuples into SortTuple slots during merge. Instead, use the
memory for larger read buffers in logtape.c. We're doing the same number
of READTUP() calls either way, but managing the pre-read SortTuple slots
is much more complicated. Also, the on-tape representation is more compact
than SortTuples, so we can fit more pre-read tuples into the same amount
of memory this way. And we have better cache-locality, when we use just a
small number of SortTuple slots.
Now that we only hold one tuple from each tape in the SortTuple slots, we
can greatly simplify the "batch memory" management. We now maintain a
small set of fixed-sized slots, to hold the tuples, and fall back to
palloc() for larger tuples. We use this method during all merge phases,
not just the final merge, and also when randomAccess is requested, and
also in the TSS_SORTEDONTAPE case. In other words, it's used whenever we
do an external sort.
Reviewed by Peter Geoghegan and Claudio Freire.
Discussion: <CAM3SWZTpaORV=yQGVCG8Q4axcZ3MvF-05xe39ZvORdU9JcD6hQ@mail.gmail.com>
Without this, an extension containing an access method is not properly
dumped/restored during pg_upgrade --- the AM ends up not being a member
of the extension after upgrading.
Another oversight in commit 473b93287, reported by Andrew Dunstan.
Report: <f7ac29f3-515c-2a44-21c5-ec925053265f@dunslane.net>
Fixes errors introduced in commit bc34223bc, as detected by Coverity.
In passing, report ENOSPC for a short write while padding a new wal file in
open_walfile, make certain that close_walfile closes walfile in all cases,
and improve a couple of comments.
Michael Paquier and Tom Lane
In standard Unix builds, postmaster child processes do ClosePostmasterPorts
immediately after InitPostmasterChild, that is almost immediately after
being spawned. This is important because we don't want children holding
open the postmaster's end of the postmaster death watch pipe.
However, in EXEC_BACKEND builds, SubPostmasterMain was postponing this
responsibility significantly, in order to make it slightly more convenient
to pass the right flag value to ClosePostmasterPorts. This is bad,
particularly seeing that process_shared_preload_libraries() might invoke
nearly-arbitrary code. Rearrange so that we do it as soon as we've
fetched the socket FDs via read_backend_variables().
Also move the comment explaining about randomize_va_space to before the
call of PGSharedMemoryReAttach, which is where it's relevant. The old
placement was appropriate when the reattach happened inside
CreateSharedMemoryAndSemaphores, but that was a long time ago.
Back-patch to 9.3; the patch doesn't apply cleanly before that, and
it doesn't seem worth a lot of effort given that we've had no actual
field complaints traceable to this.
Discussion: <4157.1475178360@sss.pgh.pa.us>
There is no need for "all: all-lib" to be placed before inclusion of
Makefile.shlib. Makefile.global is what ensures that "all" is the
default target, and we already document that that has to be included
first. Per comment from Pavel Raiskup.
Discussion: <1925924.izSMJEZO3x@unused-4-107.brq.redhat.com>
The sequence "configure; cd src/pl/plpython; make -j" failed due to
trying to compile plpython's .o files before the generated headers
finished building. (This is an important real-world case, since it's
the typical second step when building both plpython2 and plpython3.)
This happens because the submake-generated-headers target is not
placed in a way to make it a prerequisite to compiling the .o files.
Fix that.
Checking other uses of submake-generated-headers, I noted that the one
attached to pg_regress was similarly misplaced; but it's actually not
needed at all for pg_regress.o, rather regress.o, so move it to be a
prerequisite of that.
Back-patch to 9.6 where submake-generated-headers was introduced
(by commit 548af97fc). It's not immediately clear to me why the
previous coding didn't have the same issue; but since we've not
had field reports of plpython make failing, leave it alone in the
older branches.
Pavel Raiskup and Tom Lane
Discussion: <1925924.izSMJEZO3x@unused-4-107.brq.redhat.com>
Before pg_regress runs psql, set the application name to the test name.
Similarly, set the application name to the test file name in the TAP
tests. Also, set a default log_line_prefix that show the application
name, as well as the PID and a time stamp.
That way, the server log output can be correlated to the test input
files, making debugging a bit easier.
The previous design for this had copyFile(), linkFile(), and
rewriteVisibilityMap() returning strerror strings, with the caller
producing one-size-fits-all error messages based on that. This made it
impossible to produce messages that described the failures with any degree
of precision, especially not short-read problems since those don't set
errno at all.
Since pg_upgrade has no intention of continuing after any error in this
area, let's fix this by just letting these functions call pg_fatal() for
themselves, making it easy for each point of failure to have a suitable
error message. Taking this approach also allows dropping cleanup code
that was unnecessary and was often rather sloppy about preserving errno.
To not lose relevant info that was reported before, pass in the schema name
and table name of the current table so that they can be included in the
error reports.
An additional problem was the use of getErrorText(), which was flat out
wrong for all but a couple of call sites, because it unconditionally did
"_dosmaperr(GetLastError())" on Windows. That's only appropriate when
reporting an error from a Windows-native API, which only a couple of
the callers were actually doing. Thus, even the reported strerror string
would be unrelated to the actual failure in many cases on Windows.
To fix, get rid of getErrorText() altogether, and just have call sites
do strerror(errno) instead, since that's the way all the rest of our
frontend programs do it. Add back the _dosmaperr() calls in the two
places where that's actually appropriate.
In passing, make assorted messages hew more closely to project style
guidelines, notably by removing initial capitals in not-complete-sentence
primary error messages. (I didn't make any effort to clean up places
I didn't have another reason to touch, though.)
Per discussion of a report from Thomas Kellerer. Back-patch to 9.6,
but no further; given the relative infrequency of reports of problems
here, it's not clear it's worth adapting the patch to older branches.
Patch by me, but with credit to Alvaro Herrera for spotting the issue
with getErrorText's misuse of _dosmaperr().
Discussion: <nsjrbh$8li$1@blaine.gmane.org>
This is new code in 9.6, and evidently we missed out testing it as
thoroughly as it should have been. Bugs fixed here:
1. Use binary not text mode to open the files on Windows. Before, if
the visibility map chanced to contain two bytes that looked like \r\n,
Windows' read() would convert that to \n, which both corrupts the map
data and causes the file to look shorter than it should. Unless you
were *very* unlucky and had an exact multiple of 8K such occurrences
in each VM file, this would cause pg_upgrade to report a failure,
though with a rather obscure error message.
2. The code for copying rebuilt bytes into the output was simply wrong.
It chanced to work okay on little-endian machines but would emit the
bytes in the wrong order on big-endian, leading to silent corruption
of the visibility map data.
3. The code was careless about alignment of the working buffers. Given
all three of an alignment-picky architecture, a compiler that chooses
to put the new_vmbuf[] local variable at an odd starting address, and
a checksum-enabled database, pg_upgrade would dump core.
Point one was reported by Thomas Kellerer, the other two detected by
code-reading.
Point two is much the nastiest of these issues from an impact standpoint,
though fortunately it affects only a minority of users. The Windows issue
will definitely bite people, but it seems quite unlikely that there would
be undetected corruption from that.
In addition, I failed to resist the temptation to do some minor cosmetic
adjustments, mostly improving the comments.
It would be a good idea to try to improve the error reporting here, but
that seems like material for a separate patch.
Discussion: <nsjrbh$8li$1@blaine.gmane.org>
There is a small window between when the server closes out the existing
segment and the new one is created. Put a loop around the open call in
this case to make sure we wait for the new file to actually appear.
Several places weren't careful about fsyncing in the way. See 1d4a0ab1
and 606e0f98 for details about required fsyncs.
This adds a couple of functions in src/common/ that have an equivalent
in the backend: durable_rename(), fsync_parent_path()
From: Michael Paquier <michael.paquier@gmail.com>
The list of files and directories that pg_basebackup excludes from the
backup was somewhat incomplete and unorganized. Change that with having
the exclusion driven from tables. Clean up some code around it. Also
document the exclusions in more detail so that users of pg_start_backup
can make use of it as well.
The contents of these directories are now excluded from the backup:
pg_dynshmem, pg_notify, pg_serial, pg_snapshots, pg_subtrans
Also fix a bug that a pg_repl_slot or pg_stat_tmp being a symlink would
cause a corrupt tar header to be created. Now such symlinks are
included in the backup as empty directories. Bug found by Ashutosh
Sharma <ashu.coek88@gmail.com>.
From: David Steele <david@pgmasters.net>
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Add a validity flag to DCHCacheEntry and NUMCacheEntry entries, and
do not set it true until after we've parsed the supplied format string.
This allows dealing with possible errors while parsing the format
without the baroque hack that was there before (which only covered
errors within NUMDesc_prepare, anyway). We can get rid of the PG_TRY in
NUMDesc_prepare, as well as last_NUMCacheEntry and NUM_cache_remove.
(Essentially, this reverts commit ff783fbae in favor of a less fragile
solution; the problems with that approach are well illustrated by later
hacking such as 55f927a46.)
In passing, define the size of these caches as DCH_CACHE_ENTRIES not
DCH_CACHE_FIELDS + 1 (whoever thought that was a good definition?)
and likewise for the NUM cache. Also const-ify format string parameters
where convenient, and merge duplicated cache lookup logic.
This is primarily driven by a proposed patch from Artur Zakirov,
which introduced some ereport's into format string parsing for
the datetime case. He proposed preventing the creation of invalid
cache entries by parsing the format string first into a local-variable
array, and then copying that to a cache entry. That seemed a bit
ugly to me, and anyway randomly different from the way the identical
problem had been solved for the numeric case. Let's make the two
sets of code more similar not less so.
I'm not sure whether we'll adopt the new error conditions Artur proposes,
but this patch seems like good code cleanup and future-proofing in any
case. The existing code is critically (and undocumented-ly) dependent on
no elog being thrown out of several nontrivial functions, which is trouble
waiting to happen, though it doesn't seem to be actively broken today.
Discussion: <b2a39359-3282-b402-f4a3-057aae500ee7@postgrespro.ru>
Historically, something like to_date('2009-06-40','YYYY-MM-DD') would
return '2009-07-10' because there was no prohibition on out-of-range
month or day numbers. This has been widely panned, and it also turns
out that Oracle throws an error in such cases. Since these functions
are nominally Oracle-compatibility features, let's change that.
There's no particular restriction on year (modulo the fact that the
scanner may not believe that more than 4 digits are year digits,
a matter to be addressed separately if at all). But we now check month,
day, hour, minute, second, and fractional-second fields, as well as
day-of-year and second-of-day fields if those are used.
Currently, no checks are made on ISO-8601-style week numbers or day
numbers; it's not very clear what the appropriate rules would be there,
and they're probably so little used that it's not worth sweating over.
Artur Zakirov, reviewed by Amul Sul, further adjustments by me
Discussion: <1873520224.1784572.1465833145330.JavaMail.yahoo@mail.yahoo.com>
See-Also: <57786490.9010201@wars-nicht.de>
The previous patch broke this by returning NULL for a failed CRC check,
which pg_controldata would then try to read. Fix by returning the
result of the CRC check in a separate argument.
Michael Paquier and myself
Andres Freund