exit() is not async-signal safe. Even if the libc implementation is, 3rd
party libraries might have installed unsafe atexit() callbacks. After
receiving SIGQUIT, we really just want to exit as quickly as possible, so
we don't really want to run the atexit() callbacks anyway.
The original report by Jimmy Yih was a self-deadlock in startup_die().
However, this patch doesn't address that scenario; the signal handling
while waiting for the startup packet is more complicated. But at least this
alleviates similar problems in the SIGQUIT handlers, like that reported
by Asim R P later in the same thread.
Backpatch to 9.3 (all supported versions).
Discussion: https://www.postgresql.org/message-id/CAOMx_OAuRUHiAuCg2YgicZLzPVv5d9_H4KrL_OFsFP%3DVPekigA%40mail.gmail.com
6cb3372 enforces errno to ENOSPC when less bytes than what is expected
have been written when it is unset, though it forgot to properly reset
errno before doing a system call to write(), causing errno to
potentially come from a previous system call.
Reported-by: Tom Lane
Author: Michael Paquier
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/31797.1533326676@sss.pgh.pa.us
This was broken in commit 9c7d06d606, which inadvertently gave the
wrong value to fast_forward in one StartupDecodingContext call. Fix by
flipping the value. Add a test for the obvious error, namely trying to
initialize a replication slot with an nonexistent output plugin.
While at it, move the CreateDecodingContext call earlier, so that any
errors are reported before sending the CopyBoth message.
Author: Dave Cramer <davecramer@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CADK3HHLVkeRe1v4P02-5hj55H3_yJg3AEtpXyEY5T3wuzO2jSg@mail.gmail.com
Those would use the default ERRCODE_INTERNAL_ERROR, but for foreseeable
failures an errcode ought to be set, ERRCODE_DATA_CORRUPTED making the
most sense here.
While on the way, fix one errcode_for_file_access missing in origin.c
since the code has been created, and remove one assignment of errno to 0
before calling read(), as this was around to fit with what was present
before 811b6e36 where errno would not be set when not enough bytes are
read. I have noticed the first one, and Tom has pinged me about the
second one.
Author: Michael Paquier
Reported-by: Tom Lane
Discussion: https://postgr.es/m/27265.1531925836@sss.pgh.pa.us
As benchmarks show, using libc's string-to-integer conversion is
pretty slow. At least part of the reason for that is that strtol[l]
have to be more generic than what largely is required inside pg.
This patch considerably speeds up int2/int4 input (int8 already was
already using hand-rolled code).
Most of the existing pg_atoi callers have been converted. But as one
requires pg_atoi's custom delimiter functionality, and as it seems
likely that there's external pg_atoi users, it seems sensible to just
keep pg_atoi around.
Author: Andres Freund
Reviewed-By: Robert Haas
Discussion: https://postgr.es/m/20171208214437.qgn6zdltyq5hmjpk@alap3.anarazel.de
The code added by 9c7d06d606 was a bit obscure; clarify that by
rewriting the comments. Lack of clarity has already caused bugs, so
it's a worthy goal.
Co-authored-by: Arseny Sher <a.sher@postgrespro.ru>
Co-authored-by: Michaël Paquier <michael@paquier.xyz>
Co-authored-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Petr Jelínek <petr.jelinek@2ndquadrant.com>
Discussion: https://postgr.es/m/87y3fgoyrn.fsf@ars-thinkpad
Historically, we've allowed auxiliary processes to take buffer pins without
tracking them in a ResourceOwner. However, that creates problems for error
recovery. In particular, we've seen multiple reports of assertion crashes
in the startup process when it gets an error while holding a buffer pin,
as for example if it gets ENOSPC during a write. In a non-assert build,
the process would simply exit without releasing the pin at all. We've
gotten away with that so far just because a failure exit of the startup
process translates to a database crash anyhow; but any similar behavior
in other aux processes could result in stuck pins and subsequent problems
in vacuum.
To improve this, institute a policy that we must *always* have a resowner
backing any attempt to pin a buffer, which we can enforce just by removing
the previous special-case code in resowner.c. Add infrastructure to make
it easy to create a process-lifespan AuxProcessResourceOwner and clear
out its contents at appropriate times. Replace existing ad-hoc resowner
management in bgwriter.c and other aux processes with that. (Thus, while
the startup process gains a resowner where it had none at all before, some
other aux process types are replacing an ad-hoc resowner with this code.)
Also use the AuxProcessResourceOwner to manage buffer pins taken during
StartupXLOG and ShutdownXLOG, even when those are being run in a bootstrap
process or a standalone backend rather than a true auxiliary process.
In passing, remove some other ad-hoc resource owner creations that had
gotten cargo-culted into various other places. As far as I can tell
that was all unnecessary, and if it had been necessary it was incomplete,
due to lacking any provision for clearing those resowners later.
(Also worth noting in this connection is that a process that hasn't called
InitBufferPoolBackend has no business accessing buffers; so there's more
to do than just add the resowner if we want to touch buffers in processes
not covered by this patch.)
Although this fixes a very old bug, no back-patch, because there's no
evidence of any significant problem in non-assert builds.
Patch by me, pursuant to a report from Justin Pryzby. Thanks to
Robert Haas and Kyotaro Horiguchi for reviews.
Discussion: https://postgr.es/m/20180627233939.GA10276@telsasoft.com
A collection of typos I happened to spot while reading code, as well as
grepping for common mistakes.
Backpatch to all supported versions, as applicable, to avoid conflicts
when backporting other commits in the future.
This should tame the beast, as there are no other places where off_t is
used in the new error messages.
Reported again by longfin, which complained about walsender.c while I
spotted the other two ones while double-checking.
Some error messages related to file handling are using the code path
context to define their state. For example, 2PC-related errors are
referring to "two-phase status files", or "relation mapping file" is
used for catalog-to-filenode mapping, however those prove to be
difficult to translate, and are not more helpful than just referring to
the path of the file being worked on. So simplify all those error
messages by just referring to files with their path used. In some
cases, like the manipulation of WAL segments, the context is actually
helpful so those are kept.
Calls to the system function read() have also been rather inconsistent
with their error handling sometimes not reporting the number of bytes
read, and some other code paths trying to use an errno which has not
been set. The in-core functions are using a more consistent pattern
with this patch, which checks for both errno if set or if an
inconsistent read is happening.
So as to care about pluralization when reading an unexpected number of
byte(s), "could not read: read %d of %zu" is used as error message, with
%d field being the output result of read() and %zu the expected size.
This simplifies the work of translators with less variations of the same
message.
Author: Michael Paquier
Reviewed-by: Álvaro Herrera
Discussion: https://postgr.es/m/20180520000522.GB1603@paquier.xyz
Since the old logic was completely unaware of subtransactions, a
change made in a subsequently-aborted subtransaction would still cause
workers to be stopped at toplevel transaction commit. Fix that by
managing a stack of worker lists rather than just one.
Amit Khandekar and Robert Haas
Discussion: http://postgr.es/m/CAJ3gD9eaG_mWqiOTA2LfAug-VRNn1hrhf50Xi1YroxL37QkZNg@mail.gmail.com
All attributes and arguments using a slot name map to the data type
"name", but this function has been using "text". This is cosmetic, as
even if text is used then the slot name would be truncated to 64
characters anyway and stored as such. The documentation already said
so and the function already assumed that the argument was of this type
when fetching its value.
Bump catalog version.
Author: Sawada Masahiko
Discussion: https://postgr.es/m/CAD21AoADYz_-eAqH5AVFaCaojcRgwpo9PW=u8kgTMys63oB8Cw@mail.gmail.com
WAL senders sending logically-decoded data fail to properly report in
"streaming" state when starting up, hence as long as one extra record is
not replayed, such WAL senders would remain in a "catchup" state, which
is inconsistent with the physical cousin.
This can be easily reproduced by for example using pg_recvlogical and
restarting the upstream server. The TAP tests have been slightly
modified to detect the failure and strengthened so as future tests also
make sure that a node is in streaming state when waiting for its
catchup.
Backpatch down to 9.4 where this code has been introduced.
Reported-by: Sawada Masahiko
Author: Simon Riggs, Sawada Masahiko
Reviewed-by: Petr Jelinek, Michael Paquier, Vaishnavi Prabakaran
Discussion: https://postgr.es/m/CAD21AoB2ZbCCqOx=bgKMcLrAvs1V0ZMqzs7wBTuDySezTGtMZA@mail.gmail.com
Such replication slots are physical slots freshly created without WAL
being reserved, which is the default behavior, which have not been used
yet as WAL consumption resources to retain WAL. This prevents advancing
a slot to a position older than any WAL available, which could falsify
calculations for WAL segment recycling.
This also cleans up a bit the code, as ReplicationSlotRelease() would be
called on ERROR, and improves error messages.
Reported-by: Kyotaro Horiguchi
Author: Michael Paquier
Reviewed-by: Andres Freund, Álvaro Herrera, Kyotaro Horiguchi
Discussion: https://postgr.es/m/20180626071305.GH31353@paquier.xyz
Starting with commit 9915de6c1c, replication slot drop uses a
condition variable sleep to wait until the current user of the slot goes
away. This is more user friendly than the previous behavior of erroring
out if the slot is in use, but it fails with a not-for-user-consumption
error message in single-user mode; plus, if you're using single-user
mode because you don't want to start the server in the regular mode
(say, disk is full and WAL won't recycle because of the slot), it's
inconvenient.
Fix by skipping the cond variable sleep in single-user mode, since
there can't be anybody to wait for anyway.
Reported-by: tushar <tushar.ahuja@enterprisedb.com>
Author: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/3b2f809f-326c-38dd-7a9e-897f957a4eb1@enterprisedb.com
Coverity complains that there is no protection in the code (at least in
non-assertion-enabled builds) against speculative insertion failing to
follow the expected protocol. Add an elog(ERROR) for the case.
Two closely related bugs are fixed. First, xmin of logical slots was
advanced too early. During xl_running_xacts processing, xmin of the
slot was set to the oldest running xid in the record, but that's wrong:
actually, snapshots which will be used for not-yet-replayed transactions
might consider older txns as running too, so we need to keep xmin back
for them. The problem wasn't noticed earlier because DDL which allows
to delete tuple (set xmax) while some another not-yet-committed
transaction looks at it is pretty rare, if not unique: e.g. all forms of
ALTER TABLE which change schema acquire ACCESS EXCLUSIVE lock
conflicting with any inserts. The included test case (test_decoding's
oldest_xmin) uses ALTER of a composite type, which doesn't have such
interlocking.
To deal with this, we must be able to quickly retrieve oldest xmin
(oldest running xid among all assigned snapshots) from ReorderBuffer. To
fix, add another list of ReorderBufferTXNs to the reorderbuffer, where
transactions are sorted by base-snapshot-LSN. This is slightly
different from the existing (sorted by first-LSN) list, because a
transaction can have an earlier LSN but a later Xmin, if its first
record does not obtain an xmin (eg. xl_xact_assignment). Note this new
list doesn't fully replace the existing txn list: we still need that one
to prevent WAL recycling.
The second issue concerns SnapBuilder snapshots and subtransactions.
SnapBuildDistributeNewCatalogSnapshot never assigned a snapshot to a
transaction that is known to be a subtxn, which is good in the common
case that the top-level transaction already has one (no point in doing
so), but a bug otherwise. To fix, arrange to transfer the snapshot from
the subtxn to its top-level txn as soon as the kinship gets known.
test_decoding's snapshot_transfer verifies this.
Also, fix a minor memory leak: refcount of toplevel's old base snapshot
was not decremented when the snapshot is transferred from child.
Liberally sprinkle code comments, and rewrite a few existing ones. This
part is my (Álvaro's) contribution to this commit, as I had to write all
those comments in order to understand the existing code and Arseny's
patch.
Reported-by: Arseny Sher <a.sher@postgrespro.ru>
Diagnosed-by: Arseny Sher <a.sher@postgrespro.ru>
Co-authored-by: Arseny Sher <a.sher@postgrespro.ru>
Co-authored-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Antonin Houska <ah@cybertec.at>
Discussion: https://postgr.es/m/87lgdyz1wj.fsf@ars-thinkpad
System calls mixed up in error code paths are causing two issues which
several code paths have not correctly handled:
1) For write() calls, sometimes the system may return less bytes than
what has been written without errno being set. Some paths were careful
enough to consider that case, and assumed that errno should be set to
ENOSPC, other calls missed that.
2) errno generated by a system call is overwritten by other system calls
which may succeed once an error code path is taken, causing what is
reported to the user to be incorrect.
This patch uses the brute-force approach of correcting all those code
paths. Some refactoring could happen in the future, but this is let as
future work, which is not targeted for back-branches anyway.
Author: Michael Paquier
Reviewed-by: Ashutosh Sharma
Discussion: https://postgr.es/m/20180622061535.GD5215@paquier.xyz
Calling an external function while a pin-lock is held is a bad idea as
those are designed to be short-lived. The stress of a first commit into
a large git history may contribute to that.
Reported-by: Andres Freund
Discussion: https://postgr.es/m/20180611164952.vmxdpdpirdtkdsz6@alap3.anarazel.de
A review of the code has showed up a couple of issues fixed by this
commit:
- Physical slots have been using the confirmed LSN position as a start
comparison point which is always 0/0, instead use the restart LSN
position (logical slots need to use the confirmed LSN position, which
was correct).
- The actual slot update was incorrect for both physical and logical
slots. Physical slots need to use their restart_lsn as base comparison
point (confirmed_flush was used because of previous point), and logical
slots need to begin reading WAL from restart_lsn (confirmed_flush was
used as well), while confirmed_flush is compiled depending on the
decoding context and record read, and is the LSN position returned back
to the caller.
- Never return 0/0 if a slot cannot be advanced. This way, if a slot is
advanced while the activity is idle, then the same position is returned
to the caller over and over without raising an error. Instead return
the LSN the slot has been advanced to. With repetitive calls, the same
position is returned hence caller can directly monitor the difference in
progress in bytes by doing simply LSN difference calculations, which
should be monotonic.
Note that as the slot is owned by the backend advancing it, then the
read of those fields is fine lock-less, while updates need to happen
while the slot mutex is held, so fix that on the way as well. Other
locks for in-memory data of replication slots have been already fixed
previously.
Some of those issues have been pointed out by Petr and Simon during the
patch, while I noticed some of them after looking at the code. This
also visibly takes of a recently-discovered bug causing assertion
failures which can be triggered by a two-step slot forwarding which
first advanced the slot to a WAL page boundary and secondly advanced it
to the latest position, say 'FF/FFFFFFF' to make sure that the newest
LSN is used as forward point. It would have been nice to drop a test
for that, but the set of operators working on pg_lsn limits it, so this
is left for a future exercise.
Author: Michael Paquier
Reviewed-by: Petr Jelinek, Simon Riggs
Discussion: https://postgr.es/m/CANP8+jLyS=X-CAk59BJnsxKQfjwrmKicHQykyn52Qj-Q=9GLCw@mail.gmail.com
Discussion: https://www.postgresql.org/message-id/2840048a-1184-417a-9da8-3299d207a1d7%40postgrespro.ru
While debugging issues on HEAD for the new slot forwarding feature of
Postgres 11, some monitoring of the code surrounding in-memory slot data
has proved that the lock handling may cause inconsistent data to be read
by read-only callers of slot functions, particularly
pg_get_replication_slots() which fetches data for the system view
pg_replication_slots, or modules looking directly at slot information.
The code paths involved in those problems concern logical decoding
initialization (down to 9.4) and WAL reservation for slots (new as of
10).
A set of comments documenting all the lock handlings, particularly the
dependency with LW locks for slots and the in_use flag as well as the
internal mutex lock is added, based on a suggested by Simon Riggs.
Some of the fixed code exists down to 9.4 where WAL decoding has been
introduced, but as those race conditions are really unlikely going to
happen as those concern code paths for slot and decoding creation, just
fix the problem on HEAD.
Author: Michael Paquier
Discussion: https://postgr.es/m/20180528085747.GA27845@paquier.xyz
When due to publication configuration, a TRUNCATE change ends up with
zero tables to be published, don't send the message out, just skip it.
It's not wrong, but obviously useless overhead.
Teach both base backups and pg_verify_checksums that if a page is new,
it does not have a checksum yet, so it shouldn't be verified.
Noted by Tomas Vondra, review by David Steele.
In commit 9c0a0de4c, I'd failed to notice that catalog/catalog.h
should also be considered a frontend-unsafe header, because it includes
(and needs) the full form of pg_class.h, not to mention relcache.h.
However, various frontend code was depending on it to get
TABLESPACE_VERSION_DIRECTORY, so refactoring of some sort is called for.
The cleanest answer seems to be to move TABLESPACE_VERSION_DIRECTORY,
as well as the OIDCHARS symbol, to common/relpath.h. Do that, and mop up
inclusions as necessary. (I found that quite a few current users of
catalog/catalog.h don't seem to need it at all anymore, apparently as a
result of the refactorings that created common/relpath.[hc]. And
initdb.c needed it only as a route to pg_class_d.h.)
Discussion: https://postgr.es/m/6629.1523294509@sss.pgh.pa.us
This reverts the backend sides of commit 1fde38beaa.
I have, at least for now, left the pg_verify_checksums tool in place, as
this tool can be very valuable without the rest of the patch as well,
and since it's a read-only tool that only runs when the cluster is down
it should be a lot safer.
Consolidate directory and file create permissions for tools which work
with the PG data directory by adding a new module (common/file_perm.c)
that contains variables (pg_file_create_mode, pg_dir_create_mode) and
constants to initialize them (0600 for files and 0700 for directories).
Convert mkdir() calls in the backend to MakePGDirectory() if the
original call used default permissions (always the case for regular PG
directories).
Add tests to make sure permissions in PGDATA are set correctly by the
tools which modify the PG data directory.
Authors: David Steele <david@pgmasters.net>,
Adam Brightwell <adam.brightwell@crunchydata.com>
Reviewed-By: Michael Paquier, with discussion amongst many others.
Discussion: https://postgr.es/m/ad346fe6-b23e-59f1-ecb7-0e08390ad629%40pgmasters.net
Update the built-in logical replication system to make use of the
previously added logical decoding for TRUNCATE support. Add the
required truncate callback to pgoutput and a new logical replication
protocol message.
Publications get a new attribute to determine whether to replicate
truncate actions. When updating a publication via pg_dump from an older
version, this is not set, thus preserving the previous behavior.
Author: Simon Riggs <simon@2ndquadrant.com>
Author: Marco Nenciarini <marco.nenciarini@2ndquadrant.it>
Author: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Add a new WAL record type for TRUNCATE, which is only used when
wal_level >= logical. (For physical replication, TRUNCATE is already
replicated via SMGR records.) Add new callback for logical decoding
output plugins to receive TRUNCATE actions.
Author: Simon Riggs <simon@2ndquadrant.com>
Author: Marco Nenciarini <marco.nenciarini@2ndquadrant.it>
Author: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
We don't actually need the insert-or-update logic, so it's clearer to
have separate functions for the inserting and updating.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
In case the subscription is removed before the worker is fully started,
give a specific error message instead of the generic "cache lookup"
error.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
This makes it possible to turn checksums on in a live cluster, without
the previous need for dump/reload or logical replication (and to turn it
off).
Enabling checkusm starts a background process in the form of a
launcher/worker combination that goes through the entire database and
recalculates checksums on each and every page. Only when all pages have
been checksummed are they fully enabled in the cluster. Any failure of
the process will revert to checksums off and the process has to be
started.
This adds a new WAL record that indicates the state of checksums, so
the process works across replicated clusters.
Authors: Magnus Hagander and Daniel Gustafsson
Review: Tomas Vondra, Michael Banck, Heikki Linnakangas, Andrey Borodin
THis adds a "flags" field to the BackgroundWorkerInitializeConnection()
and BackgroundWorkerInitializeConnectionByOid(). For now only one flag,
BGWORKER_BYPASS_ALLOWCONN, is defined, which allows the worker to ignore
datallowconn.
We were being careless in some places about the order of -L switches in
link command lines, such that -L switches referring to external directories
could come before those referring to directories within the build tree.
This made it possible to accidentally link a system-supplied library, for
example /usr/lib/libpq.so, in place of the one built in the build tree.
Hilarity ensued, the more so the older the system-supplied library is.
To fix, break LDFLAGS into two parts, a sub-variable LDFLAGS_INTERNAL
and the main LDFLAGS variable, both of which are "recursively expanded"
so that they can be incrementally adjusted by different makefiles.
Establish a policy that -L switches for directories in the build tree
must always be added to LDFLAGS_INTERNAL, while -L switches for external
directories must always be added to LDFLAGS. This is sufficient to
ensure a safe search order. For simplicity, we typically also put -l
switches for the respective libraries into those same variables.
(Traditional make usage would have us put -l switches into LIBS, but
cleaning that up is a project for another day, as there's no clear
need for it.)
This turns out to also require separating SHLIB_LINK into two variables,
SHLIB_LINK and SHLIB_LINK_INTERNAL, with a similar rule about which
switches go into which variable. And likewise for PG_LIBS.
Although this change might appear to affect external users of pgxs.mk,
I think it doesn't; they shouldn't have any need to touch the _INTERNAL
variables.
In passing, tweak src/common/Makefile so that the value of CPPFLAGS
recorded in pg_config lacks "-DFRONTEND" and the recorded value of
LDFLAGS lacks "-L../../../src/common". Both of those things are
mistakes, apparently introduced during prior code rearrangements,
as old versions of pg_config don't print them. In general we don't
want anything that's specific to the src/common subdirectory to
appear in those outputs.
This is certainly a bug fix, but in view of the lack of field
complaints, I'm unsure whether it's worth the risk of back-patching.
In any case it seems wise to see what the buildfarm makes of it first.
Discussion: https://postgr.es/m/25214.1522604295@sss.pgh.pa.us
When base backups are run over the replication protocol (for example
using pg_basebackup), verify the checksums of all data blocks if
checksums are enabled. If checksum failures are encountered, log them
as warnings but don't abort the backup.
This becomes the default behaviour in pg_basebackup (provided checksums
are enabled on the server), so add a switch (-k) to disable the checks
if necessary.
Author: Michael Banck
Reviewed-By: Magnus Hagander, David Steele
Discussion: https://postgr.es/m/20180228180856.GE13784@nighthawk.caipicrew.dd-dns.de
Previously there was no way in the standby side to find out the host and port
of the sender server that the walreceiver was currently connected to when
multiple hosts and ports were specified in primary_conninfo. For that purpose,
this patch adds sender_host and sender_port columns into pg_stat_wal_receiver
view. They report the host and port that the active replication connection
currently uses.
Bump catalog version.
Author: Haribabu Kommi
Reviewed-by: Michael Paquier and me
Discussion: https://postgr.es/m/CAJrrPGcV_aq8=cdqkFhVDJKEnDQ70yRTTdY9RODzMnXNrCz2Ow@mail.gmail.com
The target cluster that was rewound needs to perform recovery from
the checkpoint created at failover, which leads it to remove or recreate
some files and directories that may have been copied from the source
cluster. So pg_rewind can skip synchronizing such files and directories,
and which reduces the amount of data transferred during a rewind
without changing the usefulness of the operation.
Author: Michael Paquier
Reviewed-by: Anastasia Lubennikova, Stephen Frost and me
Discussion: https://postgr.es/m/20180205071022.GA17337@paquier.xyz
Originally, we treated memory context names as potentially variable in
all cases, and therefore always copied them into the context header.
Commit 9fa6f00b1 rethought this a little bit and invented a distinction
between fixed and variable names, skipping the copy step for the former.
But we can make things both simpler and more useful by instead allowing
there to be two parts to a context's identification, a fixed "name" and
an optional, variable "ident". The name supplied in the context create
call is now required to be a compile-time-constant string in all cases,
as it is never copied but just pointed to. The "ident" string, if
wanted, is supplied later. This is needed because typically we want
the ident to be stored inside the context so that it's cleaned up
automatically on context deletion; that means it has to be copied into
the context before we can set the pointer.
The cost of this approach is basically just an additional pointer field
in struct MemoryContextData, which isn't much overhead, and is bought
back entirely in the AllocSet case by not needing a headerSize field
anymore, since we no longer have to cope with variable header length.
In addition, we can simplify the internal interfaces for memory context
creation still further, saving a few cycles there. And it's no longer
true that a custom identifier disqualifies a context from participating
in aset.c's freelist scheme, so possibly there's some win on that end.
All the places that were using non-compile-time-constant context names
are adjusted to put the variable info into the "ident" instead. This
allows more effective identification of those contexts in many cases;
for example, subsidary contexts of relcache entries are now identified
by both type (e.g. "index info") and relname, where before you got only
one or the other. Contexts associated with PL function cache entries
are now identified more fully and uniformly, too.
I also arranged for plancache contexts to use the query source string
as their identifier. This is basically free for CachedPlanSources, as
they contained a copy of that string already. We pay an extra pstrdup
to do it for CachedPlans. That could perhaps be avoided, but it would
make things more fragile (since the CachedPlanSource is sometimes
destroyed first). I suspect future improvements in error reporting will
require CachedPlans to have a copy of that string anyway, so it's not
clear that it's worth moving mountains to avoid it now.
This also changes the APIs for context statistics routines so that the
context-specific routines no longer assume that output goes straight
to stderr, nor do they know all details of the output format. This
is useful immediately to reduce code duplication, and it also allows
for external code to do something with stats output that's different
from printing to stderr.
The reason for pushing this now rather than waiting for v12 is that
it rethinks some of the API changes made by commit 9fa6f00b1. Seems
better for extension authors to endure just one round of API changes
not two.
Discussion: https://postgr.es/m/CAB=Je-FdtmFZ9y9REHD7VsSrnCkiBhsA4mdsLKSPauwXtQBeNA@mail.gmail.com
Commit 8694cc96b did this randomly differently from other callers of
parse_filename_for_nontemp_relation(). Perhaps unsurprisingly,
the randomly different way is wrong; it fails to ensure the
extracted string is null-terminated. Per buildfarm member skink.
Discussion: https://postgr.es/m/14453.1522001792@sss.pgh.pa.us
Per the project style guide, details and hints should have leading
capitalization and end with a period. On the other hand, errcontext should
not be capitalized and should not end with a period. To support well
formatted error contexts in dblink, extend dblink_res_error() to take a
format+arguments rather than a hardcoded string.
Daniel Gustafsson
Discussion: https://postgr.es/m/B3C002C8-21A0-4F53-A06E-8CAB29FCF295@yesql.se
Logical decoding should not publish anything about tables created as
part of a heap rewrite during DDL. Those tables don't exist externally,
so consumers of logical decoding cannot do anything sensible with that
information. In ab28feae2b, we worked
around this for built-in logical replication, but that was hack.
This is a more proper fix: We mark such transient heaps using the new
field pg_class.relwrite, linking to the original relation OID. By
default, we ignore them in logical decoding before they get to the
output plugin. Optionally, a plugin can register their interest in
getting such changes, if they handle DDL specially, in which case the
new field will help them get information about the actual table.
Reviewed-by: Craig Ringer <craig@2ndquadrant.com>
We call this thing a "transaction block" everywhere except in a few
functions, where it is mysteriously called a "transaction chain". In
the SQL standard, a transaction chain is something different. So rename
these functions to match the common terminology.
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
The logical replication type map seems to have been misused by its only
caller -- it would try to use the remote OID as input for local type
routines, which unsurprisingly could result in bogus "cache lookup
failed for type XYZ" errors, or random other type names being picked up
if they happened to use the right OID. Fix that, changing
Oid logicalrep_typmap_getid(Oid remoteid) to
char *logicalrep_typmap_gettypname(Oid remoteid)
which is more useful. If the remote type is not part of the typmap,
this simply prints "unrecognized type" instead of choking trying to
figure out -- a pointless exercise (because the only input for that
comes from replication messages, which are not under the local node's
control) and dangerous to boot, when called from within an error context
callback.
Once that is done, it comes to light that the local OID in the typmap
entry was not being used for anything; the type/schema names are what we
need, so remove local type OID from that struct.
Once you do that, it becomes pointless to attach a callback to regular
syscache invalidation. So remove that also.
Reported-by: Dang Minh Huong
Author: Masahiko Sawada
Reviewed-by: Álvaro Herrera, Petr Jelínek, Dang Minh Huong, Atsushi Torikoshi
Discussion: https://postgr.es/m/75DB81BEEA95B445AE6D576A0A5C9E936A6BE964@BPXM05GP.gisp.nec.co.jp
Discussion: https://postgr.es/m/75DB81BEEA95B445AE6D576A0A5C9E936A6C4B0A@BPXM05GP.gisp.nec.co.jp
If a walsender exits leaving data in reorderbuffers, the next walsender
that tries to decode the same transaction would append its decoded data
in the same spill files without truncating it first, which effectively
duplicate the data. Avoid that by removing any leftover reorderbuffer
spill files when a walsender starts.
Backpatch to 9.4; this bug has been there from the very beginning of
logical decoding.
Author: Craig Ringer, revised by me
Reviewed by: Álvaro Herrera, Petr Jelínek, Masahiko Sawada
In the pgoutput plugin, skip changes for relations that are not
publishable, per is_publishable_class(). This concerns in particular
materialized views and information_schema tables. While those relations
cannot be part of a publication, per existing checks, they will be
considered by a FOR ALL TABLES publication. A subscription would not
actually apply changes for those relations, again per existing checks,
but trying to match incoming changes to local tables on the subscriber
would lead to errors if no matching local table exists. Skipping those
changes on the publisher avoids sending useless changes and eliminates
the error.
Bug: #15044
Reported-by: Chad Trabant <chad@iris.washington.edu>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndquadrant.com>
The reason for doing so is that it will allow expression evaluation to
optimize based on the underlying tupledesc. In particular it will
allow to JIT tuple deforming together with the expression itself.
For that expression initialization needs to be moved after the
relevant slots are initialized - mostly unproblematic, except in the
case of nodeWorktablescan.c.
After doing so there's no need for ExecAssignResultType() and
ExecAssignResultTypeFromTL() anymore, as all former callers have been
converted to create a slot with a fixed descriptor.
When creating a slot with a fixed descriptor, tts_values/isnull can be
allocated together with the main slot, reducing allocation overhead
and increasing cache density a bit.
Author: Andres Freund
Discussion: https://postgr.es/m/20171206093717.vqdxe5icqttpxs3p@alap3.anarazel.de
Other header files should never #include postgres.h (nor postgres_fe.h,
nor c.h), per project policy. Also, there's no need for any backend .c
file to explicitly include elog.h or palloc.h, because postgres.h pulls
those in already.
Extracted from a larger patch by Kyotaro Horiguchi. The rest of the
removals he suggests require more study, but these are no-brainers.
Discussion: https://postgr.es/m/20180215.200447.209320006.horiguchi.kyotaro@lab.ntt.co.jp
Ability to advance both physical and logical replication slots using a
new user function pg_replication_slot_advance().
For logical advance that means records are consumed as fast as possible
and changes are not given to output plugin for sending. Makes 2nd phase
(after we reached SNAPBUILD_FULL_SNAPSHOT) of replication slot creation
faster, especially when there are big transactions as the reorder buffer
does not have to deal with data changes and does not have to spill to
disk.
Author: Petr Jelinek
Reviewed-by: Simon Riggs
replorigin_drop() misunderstood the API for condition variables: it
had ConditionVariablePrepareToSleep and ConditionVariableCancelSleep
inside its test-and-sleep loop, rather than outside the loop as
intended. The net effect is a narrow race-condition window wherein,
if the process using a replication slot releases it immediately after
replorigin_drop() releases the ReplicationOriginLock, replorigin_drop()
would get into the condition variable's wait list too late and then
wait indefinitely for a signal that won't come.
Because there's a different CV for each replication slot, we can't
just move the ConditionVariablePrepareToSleep call to above the
test-and-sleep loop. What we can do, in the wake of commit 13db3b936,
is drop the ConditionVariablePrepareToSleep call entirely. This fix
depends on that commit because (at least in principle) the slot matching
the target replication origin might move around, so that once in a blue
moon successive loop iterations might involve different CVs. We can now
cope with such a scenario, at the cost of an extra trip through the
retry loop.
(There are ways we could fix this bug without depending on that commit,
but they're all a lot more complicated than this way.)
While at it, upgrade the rather skimpy comments in this function.
Back-patch to v10 where this code came in.
Discussion: https://postgr.es/m/19947.1515455433@sss.pgh.pa.us
25fff40798 introduced
default monitoring roles. Apply these corrections:
* Allow access to pg_stat_get_wal_senders()
by role pg_read_all_stats
* Correct comment in pg_stat_get_wal_receiver()
to show it is no longer superuser-only.
Author: Feike Steenbergen
Reviewed-by: Michael Paquier
Apply to HEAD, then later backpatch to 10
Logical decoding's reorderbuffer.c may spill transaction files to disk
when transactions are large. These are supposed to be removed when they
become "too old" by xid; but file removal requires the boundary LSNs of
the transaction to be known. The final_lsn is only set when we see the
commit or abort record for the transaction, but nothing sets the value
for transactions that crash, so the removal code misbehaves -- in
assertion-enabled builds, it crashes by a failed assertion.
To fix, modify the final_lsn of transactions that don't have a value
set, to the LSN of the very latest change in the transaction. This
causes the spilled files to be removed appropriately.
Author: Atsushi Torikoshi
Reviewed-by: Kyotaro HORIGUCHI, Craig Ringer, Masahiko Sawada
Discussion: https://postgr.es/m/54e4e488-186b-a056-6628-50628e4e4ebc@lab.ntt.co.jp
When walsenders were included in pg_stat_activity, only the ones
actually streaming WAL were listed as active when they were active. In
particular, the connections sending base backups were listed as being
idle. Which means that a regular pg_basebackup would show up with one
active and one idle connection, when both were active.
This patch updates to set all walsenders to active when they are
(including those doing very fast things like IDENTIFY_SYSTEM), and then
back to idle. Details about exactly what they are doing is available in
pg_stat_replication.
Patch by me, review by Michael Paquier and David Steele.
A momentary window exists when synchronous_standby_names
changes that allows commands issued after the change to
continue to act as async until the change becomes visible.
Remove the race by using more appropriate test in syncrep.c
Author: Asim Rama Praveen and Ashwin Agrawal
Reported-by: Xin Zhang, Ashwin Agrawal, and Asim Rama Praveen
Reviewed-by: Michael Paquier, Masahiko Sawada
Previously an assertion failure occurred when pg_stop_backup() for
non-exclusive backup was aborted while it's waiting for WAL files to
be archived. This assertion failure happened in do_pg_abort_backup()
which was called when a non-exclusive backup was canceled.
do_pg_abort_backup() assumes that there is at least one non-exclusive
backup running when it's called. But pg_stop_backup() can be canceled
even after it marks the end of non-exclusive backup (e.g.,
during waiting for WAL archiving). This broke the assumption that
do_pg_abort_backup() relies on, and which caused an assertion failure.
This commit changes do_pg_abort_backup() so that it does nothing
when non-exclusive backup has been already marked as completed.
That is, the asssumption is also changed, and do_pg_abort_backup()
now can handle even the case where it's called when there is
no running backup.
Backpatch to 9.6 where SQL-callable non-exclusive backup was added.
Author: Masahiko Sawada and Michael Paquier
Reviewed-By: Robert Haas and Fujii Masao
Discussion: https://www.postgresql.org/message-id/CAD21AoD2L1Fu2c==gnVASMyFAAaq3y-AQ2uEVj-zTCGFFjvmDg@mail.gmail.com
The logical slots have a fast code path for sending data so as not to
impose too high a per message overhead. The fast path skips checks for
interrupts and timeouts. However, the existing coding failed to consider
the fact that a transaction with a large number of changes may take a
very long time to be processed and sent to the client. This causes the
walsender to ignore interrupts for potentially a long time and more
importantly it will result in the walsender being killed due to
timeout at the end of such a transaction.
This commit changes the fast path to also check for interrupts and only
allows calling the fast path when the last keepalive check happened less
than half the walsender timeout ago. Otherwise the slower code path will
be taken.
Backpatched to 9.4
Petr Jelinek, reviewed by Kyotaro HORIGUCHI, Yura Sokolov, Craig
Ringer and Robert Haas.
Discussion: https://postgr.es/m/e082a56a-fd95-a250-3bae-0fff93832510@2ndquadrant.com
This patch makes a number of interrelated changes to reduce the overhead
involved in creating/deleting memory contexts. The key ideas are:
* Include the AllocSetContext header of an aset.c context in its first
malloc request, rather than allocating it separately in TopMemoryContext.
This means that we now always create an initial or "keeper" block in an
aset, even if it never receives any allocation requests.
* Create freelists in which we can save and recycle recently-destroyed
asets (this idea is due to Robert Haas).
* In the common case where the name of a context is a constant string,
just store a pointer to it in the context header, rather than copying
the string.
The first change eliminates a palloc/pfree cycle per context, and
also avoids bloat in TopMemoryContext, at the price that creating
a context now involves a malloc/free cycle even if the context never
receives any allocations. That would be a loser for some common
usage patterns, but recycling short-lived contexts via the freelist
eliminates that pain.
Avoiding copying constant strings not only saves strlen() and strcpy()
overhead, but is an essential part of the freelist optimization because
it makes the context header size constant. Currently we make no
attempt to use the freelist for contexts with non-constant names.
(Perhaps someday we'll need to think harder about that, but in current
usage, most contexts with custom names are long-lived anyway.)
The freelist management in this initial commit is pretty simplistic,
and we might want to refine it later --- but in common workloads that
will never matter because the freelists will never get full anyway.
To create a context with a non-constant name, one is now required to
call AllocSetContextCreateExtended and specify the MEMCONTEXT_COPY_NAME
option. AllocSetContextCreate becomes a wrapper macro, and it includes
a test that will complain about non-string-literal context name
parameters on gcc and similar compilers.
An unfortunate side effect of making AllocSetContextCreate a macro is
that one is now *required* to use the size parameter abstraction macros
(ALLOCSET_DEFAULT_SIZES and friends) with it; the pre-9.6 habit of
writing out individual size parameters no longer works unless you
switch to AllocSetContextCreateExtended.
Internally to the memory-context-related modules, the context creation
APIs are simplified, removing the rather baroque original design whereby
a context-type module called mcxt.c which then called back into the
context-type module. That saved a bit of code duplication, but not much,
and it prevented context-type modules from exercising control over the
allocation of context headers.
In passing, I converted the test-and-elog validation of aset size
parameters into Asserts to save a few more cycles. The original thought
was that callers might compute size parameters on the fly, but in practice
nobody does that, so it's useless to expend cycles on checking those
numbers in production builds.
Also, mark the memory context method-pointer structs "const",
just for cleanliness.
Discussion: https://postgr.es/m/2264.1512870796@sss.pgh.pa.us
do_pg_start_backup() expects its callers to pass in an open DIR pointer
for the pg_tblspc directory, but there's no apparent advantage in that.
It complicates the callers without adding any flexibility, and there's no
robustness advantage, since we surely have to be prepared for errors during
the scan of pg_tblspc anyway. In fact, by holding an extra kernel resource
during operations like the preliminary checkpoint, we might be making
things a fraction more failure-prone not less. Hence, remove that argument
and open the directory just for the duration of the actual scan.
Discussion: https://postgr.es/m/28752.1512413887@sss.pgh.pa.us
This patch fixes a couple of low-probability bugs that could lead to
reporting an irrelevant errno value (and hence possibly a wrong SQLSTATE)
concerning directory-open or file-open failures. It also fixes places
where we took shortcuts in reporting such errors, either by using elog
instead of ereport or by using ereport but forgetting to specify an
errcode. And it eliminates a lot of just plain redundant error-handling
code.
In service of all this, export fd.c's formerly-static function
ReadDirExtended, so that external callers can make use of the coding
pattern
dir = AllocateDir(path);
while ((de = ReadDirExtended(dir, path, LOG)) != NULL)
if they'd like to treat directory-open failures as mere LOG conditions
rather than errors. Also fix FreeDir to be a no-op if we reach it
with dir == NULL, as such a coding pattern would cause.
Then, remove code at many call sites that was throwing an error or log
message for AllocateDir failure, as ReadDir or ReadDirExtended can handle
that job just fine. Aside from being a net code savings, this gets rid of
a lot of not-quite-up-to-snuff reports, as mentioned above. (In some
places these changes result in replacing a custom error message such as
"could not open tablespace directory" with more generic wording "could not
open directory", but it was agreed that the custom wording buys little as
long as we report the directory name.) In some other call sites where we
can't just remove code, change the error reports to be fully
project-style-compliant.
Also reorder code in restoreTwoPhaseData that was acquiring a lock
between AllocateDir and ReadDir; in the unlikely but surely not
impossible case that LWLockAcquire changes errno, AllocateDir failures
would be misreported. There is no great value in opening the directory
before acquiring TwoPhaseStateLock, so just do it in the other order.
Also fix CheckXLogRemoved to guarantee that it preserves errno,
as quite a number of call sites are implicitly assuming. (Again,
it's unlikely but I think not impossible that errno could change
during a SpinLockAcquire. If so, this function was broken for its
own purposes as well as breaking callers.)
And change a few places that were using not-per-project-style messages,
such as "could not read directory" when "could not open directory" is
more correct.
Back-patch the exporting of ReadDirExtended, in case we have occasion
to back-patch some fix that makes use of it; it's not needed right now
but surely making it global is pretty harmless. Also back-patch the
restoreTwoPhaseData and CheckXLogRemoved fixes. The rest of this is
essentially cosmetic and need not get back-patched.
Michael Paquier, with a bit of additional work by me
Discussion: https://postgr.es/m/CAB7nPqRpOCxjiirHmebEFhXVTK7V5Jvw4bz82p7Oimtsm3TyZA@mail.gmail.com
Add new style of memory allocator, known as Generational
appropriate for use in cases where memory is allocated
and then freed in roughly oldest first order (FIFO).
Use new allocator for logical decoding’s reorderbuffer
to significantly reduce memory usage and improve performance.
Author: Tomas Vondra
Reviewed-by: Simon Riggs
The lower case spellings are C and C++ standard and are used in most
parts of the PostgreSQL sources. The upper case spellings are only used
in some files/modules. So standardize on the standard spellings.
The APIs for ICU, Perl, and Windows define their own TRUE and FALSE, so
those are left as is when using those APIs.
In code comments, we use the lower-case spelling for the C concepts and
keep the upper-case spelling for the SQL concepts.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Add docs to explain this for other backup mechanisms
Author: David Steele <david@pgmasters.net>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndQuadrant.com> et al
When a publisher table has fewer columns than a subscriber, the update
of a row on the publisher should result in updating of only the columns
in common. The previous coding mistakenly reset the values of
additional columns on the subscriber to NULL because it failed to skip
updates of columns not found in the attribute map.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Flex generates a lot of functions that are not actually used. In order
to avoid coverage figures being ruined by that, mark up the part of the
.l files where the generated code appears by lcov exclusion markers.
That way, lcov will typically only reported on coverage for the .l file,
which is under our control, but not for the .c file.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Commit 597a87ccc introduced a latch pointer variable to replace use
of a long-lived shared latch in the shared WalRcvData structure.
This was not well thought out, because there are now hazards of the
pointer variable changing while it's being inspected by another
process. This could obviously lead to a core dump in code like
if (WalRcv->latch)
SetLatch(WalRcv->latch);
and there's a more remote risk of a torn read, if we have any
platforms where reading/writing a pointer is not atomic.
An actual problem would occur only if the walreceiver process
exits (gracefully) while the startup process is trying to
signal it, but that seems well within the realm of possibility.
To fix, treat the pointer variable (not the referenced latch)
as being protected by the WalRcv->mutex spinlock. There
remains a race condition that we could apply SetLatch to a
process latch that no longer belongs to the walreceiver, but
I believe that's harmless: at worst it'd cause an extra wakeup
of the next process to use that PGPROC structure.
Back-patch to v10 where the faulty code was added.
Discussion: https://postgr.es/m/22735.1507048202@sss.pgh.pa.us
1. Since commit b1a9bad9e7 we had pstrdup() inside a
spinlock-protected critical section; reported by Andreas Seltenreich.
Turn those into strlcpy() to stack-allocated variables instead.
Backpatch to 9.6.
2. Since commit 9ed551e0a4 we had a pfree() uselessly inside a
spinlock-protected critical section. Tom Lane noticed in code review.
Move down. Backpatch to 9.6.
3. Since commit 64233902d2 we had GetCurrentTimestamp() (a kernel
call) inside a spinlock-protected critical section. Tom Lane noticed in
code review. Move it up. Backpatch to 9.2.
4. Since commit 1bb2558046 we did elog(PANIC) while holding spinlock.
Tom Lane noticed in code review. Release spinlock before dying.
Backpatch to 9.2.
Discussion: https://postgr.es/m/87h8vhtgj2.fsf@ansel.ydns.eu
Add bgw_type field to background worker structure. It is intended to be
set to the same value for all workers of the same type, so they can be
grouped in pg_stat_activity, for example.
The backend_type column in pg_stat_activity now shows bgw_type for a
background worker. The ps listing also no longer calls out that a
process is a background worker but just show the bgw_type. That way,
being a background worker is more of an implementation detail now that
is not shown to the user. However, most log messages still refer to
'background worker "%s"'; otherwise constructing sensible and
translatable log messages would become tricky.
Reviewed-by: Michael Paquier <michael.paquier@gmail.com>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
A FOR ALL TABLES publication naturally considers all base tables to be a
candidate for replication. This includes transient heaps that are
created during a table rewrite during DDL. This causes failures on the
subscriber side because it will not have a table like pg_temp_16386 to
receive data (and if it did, it would be the wrong table).
The prevent this problem, we filter out any tables that match this
naming pattern and match an actual table from FOR ALL TABLES
publications. This is only a heuristic, meaning that user tables that
match that naming could accidentally be omitted. A more robust solution
might require an explicit marking of such tables in pg_class somehow.
Reported-by: yxq <yxq@o2.pl>
Bug: #14785
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Petr Jelinek <petr.jelinek@2ndquadrant.com>
The file handling functions from fd.c were called with a diverse mix of
notations for the file permissions when they were opening new files.
Almost all files created by the server should have the same permissions
set. So change the API so that e.g. OpenTransientFile() automatically
uses the standard permissions set, and OpenTransientFilePerm() is a new
function that takes an explicit permissions set for the few cases where
it is needed. This also saves an unnecessary argument for call sites
that are just opening an existing file.
While we're reviewing these APIs, get rid of the FileName typedef and
use the standard const char * for the file name and mode_t for the file
mode. This makes these functions match other file handling functions
and removes an unnecessary layer of mysteriousness. We can also get rid
of a few casts that way.
Author: David Steele <david@pgmasters.net>
For performance reasons a larger segment size than the default 16MB
can be useful. A larger segment size has two main benefits: Firstly,
in setups using archiving, it makes it easier to write scripts that
can keep up with higher amounts of WAL, secondly, the WAL has to be
written and synced to disk less frequently.
But at the same time large segment size are disadvantageous for
smaller databases. So far the segment size had to be configured at
compile time, often making it unrealistic to choose one fitting to a
particularly load. Therefore change it to a initdb time setting.
This includes a breaking changes to the xlogreader.h API, which now
requires the current segment size to be configured. For that and
similar reasons a number of binaries had to be taught how to recognize
the current segment size.
Author: Beena Emerson, editorialized by Andres Freund
Reviewed-By: Andres Freund, David Steele, Kuntal Ghosh, Michael
Paquier, Peter Eisentraut, Robert Hass, Tushar Ahuja
Discussion: https://postgr.es/m/CAOG9ApEAcQ--1ieKbhFzXSQPw_YLmepaa4hNdnY5+ZULpt81Mw@mail.gmail.com
If we failed to get a background worker slot, the code just walked
away from the logicalrep-worker slot it already had, leaving that
looking like the worker is still starting up. This led to an indefinite
hang in subscription startup, as reported by Thomas Munro. We must
release the slot on failure.
Also fix a thinko: we must capture the worker slot's generation before
releasing LogicalRepWorkerLock the first time, else testing to see if
it's changed is pretty meaningless.
BTW, the CHECK_FOR_INTERRUPTS() in WaitForReplicationWorkerAttach is a
ticking time bomb, even without considering the possibility of elog(ERROR)
in one of the other functions it calls. Really, this entire business needs
a redesign with some actual thought about error recovery. But for now
I'm just band-aiding the case observed in testing.
Back-patch to v10 where this code was added.
Discussion: https://postgr.es/m/CAEepm=2bP3TBMFBArP6o20AZaRduWjMnjCjt22hSdnA-EvrtCw@mail.gmail.com
The previous error message when attempting to run a general SQL command
in a physical replication WAL sender was a bit sloppy.
Reported-by: Fujii Masao <masao.fujii@gmail.com>
Throttling for sending a base backup in walsender is broken for the case
where there is a lot of WAL traffic, because the latch used to put the
walsender to sleep is also signalled by regular WAL traffic (and each
signal causes an additional batch of data to be sent); the net effect is
that there is no or little actual throttling. This is undesirable, so
rewrite the sleep into a loop to achieve the desired effeect.
Author: Jeff Janes, small tweaks by me
Reviewed-by: Antonin Houska
Discussion: https://postgr.es/m/CAMkU=1xH6mde-yL-Eo1TKBGNd0PB1-TMxvrNvqcAkN-qr2E9mw@mail.gmail.com
Do for replication origins what the previous commit did for replication
slots: restore the original behavior of replication origin drop to raise
an error rather than blocking, because users might be depending on the
original behavior. Maintain the blocking behavior when invoked
internally from logical replication subscription handling.
Discussion: https://postgr.es/m/20170830133922.tlpo3lgfejm4n2cs@alvherre.pgsql
Commit 9915de6c1c changed the default behavior of
DROP_REPLICATION_SLOT so that it would wait until any session holding
the slot active would release it, instead of raising an error. But
users are already depending on the original behavior, so revert to it by
default and add a WAIT option to invoke the new behavior.
Per complaint from Simone Gotti, in
Discussion: https://postgr.es/m/CAEvsy6Wgdf90O6pUvg2wSVXL2omH5OPC-38OD4Zzgk-FXavj3Q@mail.gmail.com
This is a mechanical change in preparation for a later commit that
will change the layout of TupleDesc. Introducing a macro to abstract
the details of where attributes are stored will allow us to change
that in separate step and revise it in future.
Author: Thomas Munro, editorialized by Andres Freund
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm=0ZtQ-SpsgCyzzYpsXS6e=kZWqk3g5Ygn3MDV7A8dabUA@mail.gmail.com
Change to appendStringInfoChar() or appendStringInfoString() where those
can be used.
Author: David Rowley <david.rowley@2ndquadrant.com>
Reviewed-by: Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>
Since we currently only have one protocol, this doesn't make much of a
difference other than the error message.
Author: Yugo Nagata <nagata@sraoss.co.jp>
The API for WaitLatch and friends followed the Unix convention in which
waiting for a socket connection to complete is identical to waiting for
the socket to accept a write. While Windows provides a select(2)
emulation that agrees with that, the native WaitForMultipleObjects API
treats them as quite different --- and for some bizarre reason, it will
report a not-yet-connected socket as write-ready. libpq itself has so
far escaped dealing with this because it waits with select(), but in
libpqwalreceiver.c we want to wait using WaitLatchOrSocket. The semantics
mismatch resulted in replication connection failures on Windows, but only
for remote connections (apparently, localhost connections complete
immediately, or at least too fast for anyone to have noticed the problem
in single-machine testing).
To fix, introduce an additional WL_SOCKET_CONNECTED wait flag for
WaitLatchOrSocket, which is identical to WL_SOCKET_WRITEABLE on
non-Windows, but results in waiting for FD_CONNECT events on Windows.
Ideally, we would also distinguish the two conditions in the API for
PQconnectPoll(), but changing that API at this point seems infeasible.
Instead, cheat by checking for PQstatus() == CONNECTION_STARTED to
determine that we're still waiting for the connection to complete.
(This is a cheat mainly because CONNECTION_STARTED is documented as an
internal state rather than something callers should rely on. Perhaps
we ought to change the documentation ... but this patch doesn't.)
Per reports from Jobin Augustine and Igor Neyman. Back-patch to v10
where commit 1e8a85009 exposed this longstanding shortcoming.
Andres Freund, minor fix and some code review/beautification by me
Discussion: https://postgr.es/m/CAHBggj8g2T+ZDcACZ2FmzX9CTxkWjKBsHd6NkYB4i9Ojf6K1Fw@mail.gmail.com
Similar to what was fixed in commit 9915de6c1c for replication slots,
but this time it's related to replication origins: DROP SUBSCRIPTION
attempts to drop the replication origin, but that fails if the
replication worker process hasn't yet marked it unused. This causes
failures in the buildfarm:
ERROR: could not drop replication origin with OID 1, in use by PID 34069
Like the aforementioned commit, fix by having the process running DROP
SUBSCRIPTION sleep until the worker marks the the replication origin
struct as free. This uses a condition variable on each replication
origin shmem state struct, so that the session trying to drop can sleep
and expect to be awakened by the process keeping the origin open.
Also fix a SGML markup in the previous commit.
Discussion: https://postgr.es/m/20170808001433.rozlseaf4m2wkw3n@alvherre.pgsql
In commit 9915de6c1c, we introduced a new wait point for replication
slots and incorrectly labelled it as wait event PG_WAIT_LOCK. That's
wrong, so invent an appropriate new wait event instead, and document it
properly.
While at it, fix numerous other problems in the vicinity:
- two different walreceiver wait events were being mixed up in a single
wait event (which wasn't documented either); split it out so that they
can be distinguished, and document the new events properly.
- ParallelBitmapPopulate was documented but didn't exist.
- ParallelBitmapScan was not documented (I think this should be called
"ParallelBitmapScanInit" instead.)
- Logical replication wait events weren't documented
- various symbols had been added in dartboard order in various places.
Put them in alphabetical order instead, as was originally intended.
Discussion: https://postgr.es/m/20170808181131.mu4fjepuh5m75cyq@alvherre.pgsql
This would lead to failures if local and remote tables have a different
column order. The tests previously didn't catch that because they only
tested the initial data copy. So add another test that exercises the
apply worker.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
The relation attribute map was not initialized for dropped columns,
leading to errors later on.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Scott Milliken <scott@deltaex.com>
Bug: #14769
The callers for GetOldestSafeDecodingTransactionId() all inverted the
argument for the argument introduced in 2bef06d516. Luckily this
appears to be inconsequential for the moment, as we wait for
concurrent in-progress transaction when assembling a
snapshot. Additionally this could only make a difference when adding a
second logical slot, because only a pre-existing slot could cause an
issue by lowering the returned xid dangerously much.
Reported-By: Antonin Houska
Discussion: https://postgr.es/m/32704.1496993134@localhost
Backport: 9.4-, where 2bef06d516 was backpatched to.
This fixes a crash if the local table has a function index and the
function makes non-immutable calls.
Reported-by: Scott Milliken <scott@deltaex.com>
Author: Masahiko Sawada <sawada.mshk@gmail.com>
It is relatively easy to get a replication slot to look as still active
while one process is in the process of getting rid of it; when some
other process tries to "acquire" the slot, it would fail with an error
message of "replication slot XYZ is active for PID N".
The error message in itself is fine, except that when the intention is
to drop the slot, it is unhelpful: the useful behavior would be to wait
until the slot is no longer acquired, so that the drop can proceed. To
implement this, we use a condition variable so that slot acquisition can
be told to wait on that condition variable if the slot is already
acquired, and we make any change in active_pid broadcast a signal on the
condition variable. Thus, as soon as the slot is released, the drop
will proceed properly.
Reported by: Tom Lane
Discussion: https://postgr.es/m/11904.1499039688@sss.pgh.pa.us
Authors: Petr Jelínek, Álvaro Herrera
Commit 14e8803f1 removed the locking in SyncRepWaitForLSN, but that
introduced a race condition, where SyncRepWaitForLSN might see
syncRepState already set to SYNC_REP_WAIT_COMPLETE, but the process was
not yet removed from the queue. That tripped the assertion, that the
process should no longer be in the uqeue. Reorder the operations in
SyncRepWakeQueue to remove the process from the queue first, and update
syncRepState only after that, and add a memory barrier in between to make
sure the operations are made visible to other processes in that order.
Fixes bug #14721 reported by Const Zhang. Analysis and fix by Thomas Munro.
Backpatch down to 9.5, where the locking was removed.
Discussion: https://www.postgresql.org/message-id/20170629023623.1480.26508%40wrigleys.postgresql.org
The regression tests contain numerous cases where we do some activity on a
master server and then wait till the slave has ack'd flushing its copy of
that transaction. Because WAL flush on the slave is asynchronous to the
logicalrep worker process, the worker cannot send such a feedback message
during the LogicalRepApplyLoop iteration where it processes the last data
from the master. In the previous coding, the feedback message would come
out only when the loop's WaitLatchOrSocket call returned WL_TIMEOUT. That
requires one full second of delay (NAPTIME_PER_CYCLE); and to add insult
to injury, it could take more than that if the WaitLatchOrSocket was
interrupted a few times by latch-setting events.
In reality we can expect the slave's walwriter process to have flushed the
WAL data after, more or less, WalWriterDelay (typically 200ms). Hence,
if there are unacked transactions pending, make the wait delay only that
long rather than the full NAPTIME_PER_CYCLE. Also, move one of the
send_feedback() calls into the loop main line, so that we'll check for the
need to send feedback even if we were woken by a latch event and not either
socket data or timeout.
It's not clear how much this matters for production purposes, but
it's definitely helpful for testing.
Discussion: https://postgr.es/m/30864.1498861103@sss.pgh.pa.us
When waiting for a logical replication worker process to start or stop,
we have to busy-wait until we see it add or remove itself from the
LogicalRepWorker slot in shared memory. Those loops were using a
one-second delay between checks, but on any reasonably modern machine, it
doesn't take more than a couple of msec for a worker to spawn or shut down.
Reduce the loop delays to 10ms to avoid wasting quite so much time in the
related regression tests.
In principle, a better solution would be to fix things so that the waiting
process can be awakened via its latch at the right time. But that seems
considerably more invasive, which is undesirable for a post-beta fix.
Worker start/stop performance likely isn't of huge interest anyway for
production purposes, so we might not ever get around to it.
In passing, rearrange the second wait loop in logicalrep_worker_stop()
so that the lock is held at the top of the loop, thus saving one lock
acquisition/release per call, and making it look more like the other loop.
Discussion: https://postgr.es/m/30864.1498861103@sss.pgh.pa.us
In WAL receiver and WAL server, some accesses to their corresponding
shared memory control structs were done without holding any kind of
lock, which could lead to inconsistent and possibly insecure results.
In walsender, fix by clarifying the locking rules and following them
correctly, as documented in the new comment in walsender_private.h;
namely that some members can be read in walsender itself without a lock,
because the only writes occur in the same process. The rest of the
struct requires spinlock for accesses, as usual.
In walreceiver, fix by always holding spinlock while accessing the
struct.
While there is potentially a problem in all branches, it is minor in
stable ones. This only became a real problem in pg10 because of quorum
commit in synchronous replication (commit 3901fd70cc), and a potential
security problem in walreceiver because a superuser() check was removed
by default monitoring roles (commit 25fff40798). Thus, no backpatch.
In passing, clean up some leftover braces which were used to create
unconditional blocks. Once upon a time these were used for
volatile-izing accesses to those shmem structs, which is no longer
required. Many other occurrences of this pattern remain.
Author: Michaël Paquier
Reported-by: Michaël Paquier
Reviewed-by: Masahiko Sawada, Kyotaro Horiguchi, Thomas Munro,
Robert Haas
Discussion: https://postgr.es/m/CAB7nPqTWYqtzD=LN_oDaf9r-hAjUEPAy0B9yRkhcsLdRN8fzrw@mail.gmail.com
When a sync worker is waiting for the associated apply worker to notice
that it's in SYNCWAIT state, wait_for_worker_state_change() would just
patiently wait for that to happen. This generally required waiting for
the 1-second timeout in LogicalRepApplyLoop to elapse. Kicking the worker
via its latch makes things significantly snappier.
While at it, fix race conditions that could potentially result in crashes:
we can *not* call logicalrep_worker_wakeup_ptr() once we've released the
LogicalRepWorkerLock, because worker->proc might've been reset to NULL
after we do that (indeed, there's no really solid reason to believe that
the LogicalRepWorker slot even belongs to the same worker anymore).
In logicalrep_worker_wakeup(), we can just move the wakeup inside the
lock scope. In process_syncing_tables_for_apply(), a bit more code
rearrangement is needed.
Also improve some nearby comments.
It's possible for WalSndWaitForWal to be asked to wait for WAL that doesn't
exist yet. That's fine, in fact it's the normal situation if we're caught
up; but when the client requests shutdown we should not keep waiting.
The previous coding could wait indefinitely if the source server was idle.
In passing, improve the rather weak comments in this area, and slightly
rearrange some related code for better readability.
Back-patch to 9.4 where this code was introduced.
Discussion: https://postgr.es/m/14154.1498781234@sss.pgh.pa.us
After sitting idle and fully replayed for a while and then encountering
a new burst of WAL activity, we interpolate between an ancient sample and the
not-yet-reached one for the new traffic. That produced a corner case report
of lag after receiving first new reply from standby, which might sometimes
be a large spike.
Correct this by resetting last_read time and handle that new case.
Author: Thomas Munro
Don't move parenthesized lines to the left, even if that means they
flow past the right margin.
By default, BSD indent lines up statement continuation lines that are
within parentheses so that they start just to the right of the preceding
left parenthesis. However, traditionally, if that resulted in the
continuation line extending to the right of the desired right margin,
then indent would push it left just far enough to not overrun the margin,
if it could do so without making the continuation line start to the left of
the current statement indent. That makes for a weird mix of indentations
unless one has been completely rigid about never violating the 80-column
limit.
This behavior has been pretty universally panned by Postgres developers.
Hence, disable it with indent's new -lpl switch, so that parenthesized
lines are always lined up with the preceding left paren.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
Change pg_bsd_indent to follow upstream rules for placement of comments
to the right of code, and remove pgindent hack that caused comments
following #endif to not obey the general rule.
Commit e3860ffa4d wasn't actually using
the published version of pg_bsd_indent, but a hacked-up version that
tried to minimize the amount of movement of comments to the right of
code. The situation of interest is where such a comment has to be
moved to the right of its default placement at column 33 because there's
code there. BSD indent has always moved right in units of tab stops
in such cases --- but in the previous incarnation, indent was working
in 8-space tab stops, while now it knows we use 4-space tabs. So the
net result is that in about half the cases, such comments are placed
one tab stop left of before. This is better all around: it leaves
more room on the line for comment text, and it means that in such
cases the comment uniformly starts at the next 4-space tab stop after
the code, rather than sometimes one and sometimes two tabs after.
Also, ensure that comments following #endif are indented the same
as comments following other preprocessor commands such as #else.
That inconsistency turns out to have been self-inflicted damage
from a poorly-thought-through post-indent "fixup" in pgindent.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
The new indent version includes numerous fixes thanks to Piotr Stefaniak.
The main changes visible in this commit are:
* Nicer formatting of function-pointer declarations.
* No longer unexpectedly removes spaces in expressions using casts,
sizeof, or offsetof.
* No longer wants to add a space in "struct structname *varname", as
well as some similar cases for const- or volatile-qualified pointers.
* Declarations using PG_USED_FOR_ASSERTS_ONLY are formatted more nicely.
* Fixes bug where comments following declarations were sometimes placed
with no space separating them from the code.
* Fixes some odd decisions for comments following case labels.
* Fixes some cases where comments following code were indented to less
than the expected column 33.
On the less good side, it now tends to put more whitespace around typedef
names that are not listed in typedefs.list. This might encourage us to
put more effort into typedef name collection; it's not really a bug in
indent itself.
There are more changes coming after this round, having to do with comment
indentation and alignment of lines appearing within parentheses. I wanted
to limit the size of the diffs to something that could be reviewed without
one's eyes completely glazing over, so it seemed better to split up the
changes as much as practical.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
When, during logical decoding, a transaction gets too big, it's
contents get spilled to disk. Not just the top-transaction gets
spilled, but *also* all of its subtransactions, even if they're not
that large themselves. Unfortunately we didn't clean up
such small spilled subtransactions from disk.
Fix that, by keeping better track of whether a transaction has been
spilled to disk.
Author: Andres Freund
Reported-By: Dmitriy Sarafannikov, Fabrízio de Royes Mello
Discussion:
https://postgr.es/m/1457621358.355011041@f382.i.mail.ruhttps://postgr.es/m/CAFcNs+qNMhNYii4nxpO6gqsndiyxNDYV0S=JNq0v_sEE+9PHXg@mail.gmail.com
Backpatch: 9.4-, where logical decoding was introduced
We had three occurrences of essentially the same coding pattern
wherein we tried to retrieve a query result from a libpq connection
without blocking. In the case where PQconsumeInput failed (typically
indicating a lost connection), all three loops simply gave up and
returned, forgetting to clear any previously-collected PGresult
object. Since those are malloc'd not palloc'd, the oversight results
in a process-lifespan memory leak.
One instance, in libpqwalreceiver, is of little significance because
the walreceiver process would just quit anyway if its connection fails.
But we might as well fix it.
The other two instances, in postgres_fdw, are somewhat more worrisome
because at least in principle the scenario could be repeated, allowing
the amount of memory leaked to build up to something worth worrying
about. Moreover, in these cases the loops contain CHECK_FOR_INTERRUPTS
calls, as well as other calls that could potentially elog(ERROR),
providing another way to exit without having cleared the PGresult.
Here we need to add PG_TRY logic similar to what exists in quite a
few other places in postgres_fdw.
Coverity noted the libpqwalreceiver bug; I found the other two cases
by checking all calls of PQconsumeInput.
Back-patch to all supported versions as appropriate (9.2 lacks
postgres_fdw, so this is really quite unexciting for that branch).
Discussion: https://postgr.es/m/22620.1497486981@sss.pgh.pa.us
Previously we required every exported transaction to have an xid
assigned. That was used to check that the exporting transaction is
still running, which in turn is needed to guarantee that that
necessary rows haven't been removed in between exporting and importing
the snapshot.
The exported xid caused unnecessary problems with logical decoding,
because slot creation has to wait for all concurrent xid to finish,
which in turn serializes concurrent slot creation. It also
prohibited snapshots to be exported on hot-standby replicas.
Instead export the virtual transactionid, which avoids the unnecessary
serialization and the inability to export snapshots on standbys. This
changes the file name of the exported snapshot, but since we never
documented what that one means, that seems ok.
Author: Petr Jelinek, slightly editorialized by me
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/f598b4b8-8cd7-0d54-0939-adda763d8c34@2ndquadrant.com
When a table is removed from a subscription before the tablesync worker
could start, this would previously result in an error when reading
pg_subscription_rel. Now we just ignore this.
Author: Masahiko Sawada <sawada.mshk@gmail.com>
Previously the exit handling was only able to exit from within the
main loop, and not from within the backend code it calls. Fix that by
using the standard die() SIGTERM handler, and adding the necessary
CHECK_FOR_INTERRUPTS() call.
This requires adding yet another process-type-specific branch to
ProcessInterrupts(), which hints that we probably should generalize
that handling. But that's work for another day.
Author: Petr Jelinek
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/fe072153-babd-3b5d-8052-73527a6eb657@2ndquadrant.com
Since 7c4f52409a (merged in v10), a shutdown master is reported as
FATAL: unexpected result after CommandComplete: server closed the connection unexpectedly
by walsender. It used to be
LOG: replication terminated by primary server
FATAL: could not send end-of-streaming message to primary: no COPY in progress
while the old message clearly is not perfect, it's definitely better
than what's reported now.
The change comes from the attempt to handle finished COPYs without
erroring out, needed for the new logical replication, which wasn't
needed before.
There's probably better ways to handle this, but for now just
explicitly check for a closed connection.
Author: Petr Jelinek
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/f7c7dd08-855c-e4ed-41f4-d064a6c0665a@2ndquadrant.com
Backpatch: -
A logical replication worker should not insert new rows into
pg_subscription_rel, only update existing rows, so that there are no
races if a concurrent refresh removes rows. Adjust the API to be able
to choose that behavior.
Author: Masahiko Sawada <sawada.mshk@gmail.com>
Reported-by: tushar <tushar.ahuja@enterprisedb.com>
The logical replication apply worker uses the subscription name as
application name, except for table sync. This was incorrectly set to
use the replication slot name, which might be different, in one case.
Also add a comment why the other case is different.
The larger part of this patch replaces usages of MyProc->procLatch
with MyLatch. The latter works even early during backend startup,
where MyProc->procLatch doesn't yet. While the affected code
shouldn't run in cases where it's not initialized, it might get copied
into places where it might. Using MyLatch is simpler and a bit faster
to boot, so there's little point to stick with the previous coding.
While doing so I noticed some weaknesses around newly introduced uses
of latches that could lead to missed events, and an omitted
CHECK_FOR_INTERRUPTS() call in worker_spi.
As all the actual bugs are in v10 code, there doesn't seem to be
sufficient reason to backpatch this.
Author: Andres Freund
Discussion:
https://postgr.es/m/20170606195321.sjmenrfgl2nu6j63@alap3.anarazel.dehttps://postgr.es/m/20170606210405.sim3yl6vpudhmufo@alap3.anarazel.de
Backpatch: -
Make apply busy wait check the catalog instead of shmem state to ensure
that next transaction will see the expected table synchronization state.
Also make the handover always go through same set of steps to make the
overall process easier to understand and debug.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Tested-by: Mark Kirkwood <mark.kirkwood@catalyst.net.nz>
Tested-by: Erik Rijkers <er@xs4all.nl>
This allows to cancel commands run over replication connections. While
it might have some use before v10, it has become important now that
normal SQL commands are allowed in database connected walsender
connections.
Author: Petr Jelinek
Reviewed-By: Andres Freund, Michael Paquier
Discussion: https://postgr.es/m/7966f454-7cd7-2b0c-8b70-cdca9d5a8c97@2ndquadrant.com
Because walsender and normal backends share the same main loop it's
problematic to have two different flag variables, set in signal
handlers, indicating a pending configuration reload. Only certain
walsender commands reach code paths checking for the
variable (START_[LOGICAL_]REPLICATION, CREATE_REPLICATION_SLOT
... LOGICAL, notably not base backups).
This is a bug present since the introduction of walsender, but has
gotten worse in releases since then which allow walsender to do more.
A later patch, not slated for v10, will similarly unify SIGHUP
handling in other types of processes as well.
Author: Petr Jelinek, Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170423235941.qosiuoyqprq4nu7v@alap3.anarazel.de
Backpatch: 9.2-, bug is present since 9.0
When the checkpointer writes the shutdown checkpoint, it checks
afterwards whether any WAL has been written since it started and
throws a PANIC if so. At that point, only walsenders are still
active, so one might think this could not happen, but walsenders can
also generate WAL, for instance in BASE_BACKUP and logical decoding
related commands (e.g. via hint bits). So they can trigger this panic
if such a command is run while the shutdown checkpoint is being
written.
To fix this, divide the walsender shutdown into two phases. First,
checkpointer, itself triggered by postmaster, sends a
PROCSIG_WALSND_INIT_STOPPING signal to all walsenders. If the backend
is idle or runs an SQL query this causes the backend to shutdown, if
logical replication is in progress all existing WAL records are
processed followed by a shutdown. Otherwise this causes the walsender
to switch to the "stopping" state. In this state, the walsender will
reject any further replication commands. The checkpointer begins the
shutdown checkpoint once all walsenders are confirmed as
stopping. When the shutdown checkpoint finishes, the postmaster sends
us SIGUSR2. This instructs walsender to send any outstanding WAL,
including the shutdown checkpoint record, wait for it to be replicated
to the standby, and then exit.
Author: Andres Freund, based on an earlier patch by Michael Paquier
Reported-By: Fujii Masao, Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170602002912.tqlwn4gymzlxpvs2@alap3.anarazel.de
Backpatch: 9.4, where logical decoding was introduced
The non-participation in procsignal was a problem for both changes in
master, e.g. parallelism not working for normal statements run in
walsender backends, and older branches, e.g. recovery conflicts and
catchup interrupts not working for logical decoding walsenders.
This commit thus replaces the previous WalSndXLogSendHandler with
procsignal_sigusr1_handler. In branches since db0f6cad48 that can
lead to additional SetLatch calls, but that only rarely seems to make
a difference.
Author: Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170421014030.fdzvvvbrz4nckrow@alap3.anarazel.de
Backpatch: 9.4, earlier commits don't seem to benefit sufficiently
This reverts commit 086221cf6b, which
was made to master only.
The approach implemented in the above commit has some issues. While
those could easily be fixed incrementally, doing so would make
backpatching considerably harder, so instead first revert this patch.
Discussion: https://postgr.es/m/20170602002912.tqlwn4gymzlxpvs2@alap3.anarazel.de
We didn't accept any invalidation messages until the whole sync process
had finished (because it flattens all the remote transactions in the
single one). So the sync worker didn't learn about subscription
changes/drop until it has finished. This could lead to "orphaned" sync
workers.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Masahiko Sawada <sawada.mshk@gmail.com>
This avoids "orphaned" sync workers.
This was caused by a thinko in wait_for_sync_status_change.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Masahiko Sawada <sawada.mshk@gmail.com>
When trying to access a replication slot that is supposed to already
exist, we don't need to check the naming rules again. If the slot
does not exist, we will then get a "does not exist" error message, which
is generally more useful from the perspective of an end user.
The logical replication worker processes now use the normal die()
handler for SIGTERM and CHECK_FOR_INTERRUPTS() instead of custom code.
One problem before was that the apply worker would not exit promptly
when a subscription was dropped, which could lead to deadlocks.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Masahiko Sawada <sawada.mshk@gmail.com>
Move the walrcv_disconnect() calls into the before_shmem_exit handler.
This makes sure the call is always made even during exit by signal, it
saves some duplicate code, and it makes the logic more similar to
walreceiver.c.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Logical replication supports replicating between tables with different
column order. But this failed for the initial table sync because of a
logic error in how the column list for the internal COPY command was
composed. Fix that and also add a test.
Also fix a minor omission in the column name mapping cache. When
creating the mapping list, it would not skip locally dropped columns.
So if a remote column had the same name as a locally dropped
column (...pg.dropped...), then the expected error would not occur.
Reduce some redundant messages to DEBUG1. Be clearer about the
distinction between apply workers and table synchronization workers.
Add subscription and table name where possible.
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Using flex's -i switch to achieve case-insensitivity is not a very safe
practice, because the scanner's behavior may then depend on the locale
that flex was invoked in. In the particular example at hand, that's
not academic: the possible matches for "FIRST" will be different in a
Turkish locale than elsewhere. Do it the hard way instead, as our
other scanners do.
Also, drop use of -b -CF -p, because this scanner is only used when
parsing the contents of a GUC variable. That's not done often, and
the amount of text to be parsed can be expected to be trivial, so
prioritizing scanner speed over code size seems like quite the wrong
tradeoff. Using flex's default optimization options reduces the
size of syncrep_gram.o by more than 50%.
The case-insensitivity problem is new in HEAD (cf commit 3901fd70c).
The poor choice of optimization flags exists also in 9.6, but it doesn't
seem important enough to back-patch.
Discussion: https://postgr.es/m/24403.1495225931@sss.pgh.pa.us
When creating a subscription with slot_name = NONE, we failed to check
that also create_slot = false and enabled = false were set. This
created an invalid subscription and could later lead to a crash if a
NULL slot name was accessed. Add more checks around that for
robustness.
Reported-by: tushar <tushar.ahuja@enterprisedb.com>
We used to only check for a supported relkind on the subscriber during
replication, which is needed to ensure that the setup is valid and we
don't crash. But it's also useful to tell the user immediately when
CREATE or ALTER SUBSCRIPTION is executed that the relation being added
to the subscription is not of a supported relkind.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: tushar <tushar.ahuja@enterprisedb.com>
Reformat various places in which pgindent will make a mess, and
fix a few small violations of coding style that I happened to notice
while perusing the diffs from a pgindent dry run.
There is one actual bug fix here: the need-to-enlarge-the-buffer code
path in icu_convert_case was obviously broken. Perhaps it's unreachable
in our usage? Or maybe this is just sadly undertested.
Before 955a684e04 logical decoding snapshot maintenance needed to
cope with transactions it might not have seen in their entirety. For
such transactions we'd to assume they modified the catalog (could have
happened before we were watching), and thus a new snapshot had to be
built, and distributed to concurrently running transactions.
That's problematic because building a new snapshot isn't that cheap ,
especially as the the array of committed transactions needs to be
sorted. When creating a slot on a server with a lot of transactions,
this could make logical slot creation infeasibly expensive.
After 955a684e04 there's no need to deal with transaction that
aren't guaranteed to be fully observable. That allows to avoid
building snapshots for transactions that haven't modified catalog,
even before reaching consistency.
While this isn't necessarily a bugfix, slot creation being impossible
in some production workloads, is severe enough to warrant
backpatching.
Author: Andres Freund, based on a quite different patch from Petr Jelinek
Analyzed-By: Petr Jelinek
Reviewed-By: Petr Jelinek
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backpatch: 9.4-, where logical decoding has been introduced
The snapshot assembly during the creation of logical slots relied
waiting for transactions in xl_running_xacts to end, by checking for
their commit/abort records. Unfortunately, despite locking, it is
possible to see an xl_running_xact record listing transactions as
ready, that have already WAL-logged an commit/abort record, as the
locking just prevents the ProcArray to be adjusted, and the commit
record has to be logged first.
That lead to either delayed or hanging snapshot creation, because
snapbuild.c would wait "forever" to see commit/abort records for some
transactions. That hang resolved only if a xl_running_xacts record
without any running transactions happened to be logged, far from
certain on a busy server.
It's impractical to prevent that via more heavyweight locking, the
likelihood of deadlocks and significantly increased contention would
be too big.
Instead change the initial snapshot creation to be solely based on
tracking the oldest running transaction via
xl_running_xacts->oldestRunningXid - that actually ends up
significantly simplifying the code. That has two disadvantages:
1) Because we cannot fully "trust" the contents of xl_running_xacts,
we cannot use it to build the initial snapshot. Instead we have to
wait twice for all running transactions to finish.
2) Previously a slot, unless the race occurred, could be created when
the all transaction perceived as running based on commit/abort
records, now we have to wait for the next xl_running_xacts record.
To address that, trigger logging new xl_running_xacts record from
within snapbuild.c exactly when necessary.
Unfortunately snabuild.c's SnapBuild is stored on disk, one of the
stupider ideas of a certain Mr Freund, so we can't change it in a
minor release. As this is going to be backpatched, we have to hack
around a bit to keep on-disk compatibility. A later commit will
rejigger that on master.
Author: Andres Freund, based on a quite different patch from Petr Jelinek
Analyzed-By: Petr Jelinek
Reviewed-By: Petr Jelinek
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backpatch: 9.4-, where logical decoding has been introduced
Lag tracking is called for each commit, but we introduce
a pacing delay to ensure we don't swamp the lag tracker.
Author: Petr Jelinek, with minor pacing delay code from me
Previously, the memory used by the logical replication apply worker for
processing messages would never be freed, so that could end up using a
lot of memory. To improve that, change the existing ApplyContext memory
context to ApplyMessageContext and reset that after every
message (similar to MessageContext used elsewhere). For consistency of
naming, rename the ApplyCacheContext to ApplyContext.
Author: Stas Kelvich <s.kelvich@postgrespro.ru>
It turned out this approach had problems, because a DROP command should
not have any options other than CASCADE and RESTRICT. Instead, always
attempt to drop the slot if there is one configured, but also add an
ALTER SUBSCRIPTION action to set the slot to NONE.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/29431.1493730652@sss.pgh.pa.us
This new arrangement ensures that statistics are reported right after
commit of transactions. The previous arrangement didn't get this quite
right and could lead to assertion failures.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Erik Rijkers <er@xs4all.nl>
When the checkpointer writes the shutdown checkpoint, it checks
afterwards whether any WAL has been written since it started and throws
a PANIC if so. At that point, only walsenders are still active, so one
might think this could not happen, but walsenders can also generate WAL,
for instance in BASE_BACKUP and certain variants of
CREATE_REPLICATION_SLOT. So they can trigger this panic if such a
command is run while the shutdown checkpoint is being written.
To fix this, divide the walsender shutdown into two phases. First, the
postmaster sends a SIGUSR2 signal to all walsenders. The walsenders
then put themselves into the "stopping" state. In this state, they
reject any new commands. (For simplicity, we reject all new commands,
so that in the future we do not have to track meticulously which
commands might generate WAL.) The checkpointer waits for all walsenders
to reach this state before proceeding with the shutdown checkpoint.
After the shutdown checkpoint is done, the postmaster sends
SIGINT (previously unused) to the walsenders. This triggers the
existing shutdown behavior of sending out the shutdown checkpoint record
and then terminating.
Author: Michael Paquier <michael.paquier@gmail.com>
Reported-by: Fujii Masao <masao.fujii@gmail.com>
After the logical replication launcher was told to wake up at
commit (for example, by a CREATE SUBSCRIPTION command), the flag to wake
up was not reset, so it would be woken up at every following commit as
well. So fix that by resetting the flag.
Also, we don't need to wake up anything if the transaction was rolled
back. Just reset the flag in that case.
Author: Masahiko Sawada <sawada.mshk@gmail.com>
Reported-by: Fujii Masao <masao.fujii@gmail.com>
Thinko in commit de4389712: this warning message references the wrong
"LogicalRepWorker *" variable. This would often result in a core dump,
but if it didn't, the message would show the wrong subscription OID.
In passing, adjust the message text to format a subscription OID
similarly to how that's done elsewhere in the function; and fix
grammatical issues in some nearby messages.
Per Coverity testing.
Before restarting a tablesync worker for the same relation, wait
wal_retrieve_retry_interval (currently 5s by default). This avoids
restarting failing workers in a tight loop.
We keep the last start times in a hash table last_start_times that is
separate from the table_states list, because that list is cleared out on
syscache invalidation, which happens whenever a table finishes syncing.
The hash table is kept until all tables have finished syncing.
A future project might be to unify these two and keep everything in one
data structure, but for now this is a less invasive change to accomplish
the original purpose.
For the test suite, set wal_retrieve_retry_interval to its minimum
value, to not increase the test suite run time.
Reviewed-by: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Masahiko Sawada <sawada.mshk@gmail.com>
Earlier commits (56e19d938d and 2bef06d516) make it cheaper to
create a logical slot if not exporting the initial snapshot. If
NOEXPORT_SNAPSHOT is specified, we can skip the overhead, not just
when creating a slot via sql (which can't export snapshots). As
NOEXPORT_SNAPSHOT has only recently been introduced, this shouldn't be
backpatched.
Logical decoding stores historical snapshots on disk, so that logical
decoding can restart without having to reconstruct a snapshot from
scratch (for which the resources are not guaranteed to be present
anymore). These serialized snapshots were also used when creating a
new slot via the walsender interface, which can export a "full"
snapshot (i.e. one that can read all tables, not just catalog ones).
The problem is that the serialized snapshots are only useful for
catalogs and not for normal user tables. Thus the use of such a
serialized snapshot could result in an inconsistent snapshot being
exported, which could lead to queries returning wrong data. This
would only happen if logical slots are created while another logical
slot already exists.
Author: Petr Jelinek
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backport: 9.4, where logical decoding was introduced.
Previously the logical replication launcher stored the last timestamp
when it started the worker, in the local variable "last_start_time",
in order to check whether wal_retrive_retry_interval elapsed since
the last startup of worker. If it has elapsed, the launcher sees
pg_subscription and starts new worker if necessary. This is for
limitting the startup of worker to once a wal_retrieve_retry_interval.
The bug was that the variable "last_start_time" was defined and
always initialized with 0 at the beginning of the launcher's main loop.
So even if it's set to the last timestamp in later phase of the loop,
it's always reset to 0. Therefore the launcher could not check
correctly whether wal_retrieve_retry_interval elapsed since
the last startup.
This patch moves the variable "last_start_time" outside the main loop
so that it will not be reset.
Reviewed-by: Petr Jelinek
Discussion: http://postgr.es/m/CAHGQGwGJrPO++XM4mFENAwpy1eGXKsGdguYv43GUgLgU-x8nTQ@mail.gmail.com
The logical decoding machinery already preserved all the required
catalog tuples, which is sufficient in the course of normal logical
decoding, but did not guarantee that non-catalog tuples were preserved
during computation of the initial snapshot when creating a slot over
the replication protocol.
This could cause a corrupted initial snapshot being exported. The
time window for issues is usually not terribly large, but on a busy
server it's perfectly possible to it hit it. Ongoing decoding is not
affected by this bug.
To avoid increased overhead for the SQL API, only retain additional
tuples when a logical slot is being created over the replication
protocol. To do so this commit changes the signature of
CreateInitDecodingContext(), but it seems unlikely that it's being
used in an extension, so that's probably ok.
In a drive-by fix, fix handling of
ReplicationSlotsComputeRequiredXmin's already_locked argument, which
should only apply to ProcArrayLock, not ReplicationSlotControlLock.
Reported-By: Erik Rijkers
Analyzed-By: Petr Jelinek
Author: Petr Jelinek, heavily editorialized by Andres Freund
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/9a897b86-46e1-9915-ee4c-da02e4ff6a95@2ndquadrant.com
Backport: 9.4, where logical decoding was introduced.
Publisher relation can be incorrectly chosen, if there are more than
one relation in different schemas with the same name.
Author: Euler Taveira <euler@timbira.com.br>
The code was originally written with assumption that launcher is the
only process starting the worker. However that hasn't been true since
commit 7c4f52409 which failed to modify the worker management code
adequately.
This patch adds an in_use field to the LogicalRepWorker struct to
indicate whether the worker slot is being used and uses proper locking
everywhere this flag is set or read.
However if the parent process dies while the new worker is starting and
the new worker fails to attach to shared memory, this flag would never
get cleared. We solve this rare corner case by adding a sort of garbage
collector for in_use slots. This uses another field in the
LogicalRepWorker struct named launch_time that contains the time when
the worker was started. If any request to start a new worker does not
find free slot, we'll check for workers that were supposed to start but
took too long to actually do so, and reuse their slot.
In passing also fix possible race conditions when stopping a worker that
hasn't finished starting yet.
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Fujii Masao <masao.fujii@gmail.com>
In quorum-based synchronous replication, all the standbys listed in
synchronous_standby_names equally have chances to be chosen
as synchronous standbys. So they should have the same priority.
However, previously, quorum standbys whose names appear earlier
in the list were given higher priority values though the difference of
those priority values didn't affect the selection of synchronous standbys.
Users could see those "meaningless" priority values in pg_stat_replication
and this was confusing.
This commit gives all the quorum synchronous standbys the same
highest priority, i.e., 1, in order to remove such confusion.
Author: Fujii Masao
Reviewed-by: Masahiko Sawada, Kyotaro Horiguchi
Discussion: http://postgr.es/m/CAHGQGwEKOw=SmPLxJzkBsH6wwDBgOnVz46QjHbtsiZ-d-2RGUg@mail.gmail.com
This seems to be largely cosmetic, avoiding valgrind bleats and the
like. The uninitialized padding influences the CRC of the on-disk
entry, but because it's also used when verifying the CRC, that doesn't
cause spurious failures. Backpatch nonetheless.
It's a bit unfortunate that contrib/test_decoding/sql/replorigin.sql
doesn't exercise the checkpoint path, but checkpoints are fairly
expensive on weaker machines, and we'd have to stop/start for that to
be meaningful.
Author: Andres Freund
Discussion: https://postgr.es/m/20170422183123.w2jgiuxtts7qrqaq@alap3.anarazel.de
Backpatch: 9.5, where replication origins were introduced
As reported by buildfarm animal skink / valgrind, some of the
variables weren't always initialized. To avoid further mishaps use
memset to ensure the entire entry is initialized.
Author: Petr Jelinek
Reported-By: Andres Freund
Discussion: https://postgr.es/m/20170422183123.w2jgiuxtts7qrqaq@alap3.anarazel.de
Backpatch: none, code new in master
Bug was masked by error in running 004_timeline_switch.pl that was
fixed recently in 7d68f2281a.
Detective work by Alvaro Herrera and Tom Lane
Author: Thomas Munro
Commit 7c4f524 allowed walsender to execute normal SQL commands
to support table sync feature in logical replication. Previously
while log_statement caused such SQL commands to be logged,
log_replication_commands caused them to be logged, too.
That is, such SQL commands were logged twice unexpectedly
when those settings were both enabled.
This commit forces log_replication_commands to log only replication
commands, to prevent normal SQL commands from being logged twice.
Author: Masahiko Sawada
Reviewed-by: Kyotaro Horiguchi
Reported-by: Fujii Masao
Discussion: http://postgr.es/m/CAHGQGwFDWh_Qr-q_GEMpD+qH=vYPMdVqw=ZOSY3kX_Pna9R9SA@mail.gmail.com
It's not safe to raise an error while holding spinlock. But previously
logical replication worker for table sync called the function which
reads the system catalog and may raise an error while it's holding
spinlock. Which could lead to the trouble where spinlock will never
be released and the server gets stuck infinitely.
Author: Petr Jelinek
Reviewed-by: Kyotaro Horiguchi and Fujii Masao
Reported-by: Fujii Masao
Discussion: http://postgr.es/m/CAHGQGwFDWh_Qr-q_GEMpD+qH=vYPMdVqw=ZOSY3kX_Pna9R9SA@mail.gmail.com
* Be sure to reset the launcher's pid (LogicalRepCtx->launcher_pid) to 0
even when the launcher emits an error.
* Declare ApplyLauncherWakeup() as a static function because it's called
only in launcher.c.
* Previously IsBackendPId() was used to check whether the launcher's pid
was valid. IsBackendPid() was necessary because there was the bug where
the launcher's pid was not reset to 0. But now it's fixed, so IsBackendPid()
is not necessary and this patch removes it.
Author: Masahiko Sawada
Reviewed-by: Kyotaro Horiguchi
Reported-by: Fujii Masao
Discussion: http://postgr.es/m/CAHGQGwFDWh_Qr-q_GEMpD+qH=vYPMdVqw=ZOSY3kX_Pna9R9SA@mail.gmail.com
CopyFrom() needs a range table for formatting certain errors for
constraint violations.
This changes the mechanism of how the range table is passed to the
CopyFrom() executor state. We used to generate the range table and one
entry for the relation manually inside DoCopy(). Now we use
addRangeTableEntryForRelation() to setup the range table and relation
entry for the ParseState, which is then passed down by BeginCopyFrom().
Author: Petr Jelinek <petr.jelinek@2ndquadrant.com>
Reported-by: Euler Taveira <euler@timbira.com.br>
Coverity complained because bgw.bgw_extra wasn't being filled in by
ApplyLauncherRegister(). The most future-proof fix is to memset the
whole BackgroundWorker struct to zeroes. While at it, let's apply the
same coding rule to other places that set up BackgroundWorker structs;
four out of five had the same or related issues.