If a tuple was locked by transaction A, and transaction B updated it,
the new version of the tuple created by B would be locked by A, yet
visible only to B; due to an oversight in HeapTupleSatisfiesUpdate, the
lock held by A wouldn't get checked if transaction B later deleted (or
key-updated) the new version of the tuple. This might cause referential
integrity checks to give false positives (that is, allow deletes that
should have been rejected).
This is an easy oversight to have made, because prior to improved tuple
locks in commit 0ac5ad5134 it wasn't possible to have tuples created by
our own transaction that were also locked by remote transactions, and so
locks weren't even considered in that code path.
It is recommended that foreign keys be rechecked manually in bulk after
installing this update, in case some referenced rows are missing with
some referencing row remaining.
Per bug reported by Daniel Wood in
CAPweHKe5QQ1747X2c0tA=5zf4YnS2xcvGf13Opd-1Mq24rF1cQ@mail.gmail.com
Tuple freezing was broken in connection to MultiXactIds; commit
8e53ae025d tried to fix it, but didn't go far enough. As noted by
Noah Misch, freezing a tuple whose Xmax is a multi containing an aborted
update might cause locks in the multi to go ignored by later
transactions. This is because the code depended on a multixact above
their cutoff point not having any lock-only member older than the cutoff
point for Xids, which is easily defeated in READ COMMITTED transactions.
The fix for this involves creating a new MultiXactId when necessary.
But this cannot be done during WAL replay, and moreover multixact
examination requires using CLOG access routines which are not supposed
to be used during WAL replay either; so tuple freezing cannot be done
with the old freeze WAL record. Therefore, separate the freezing
computation from its execution, and change the WAL record to carry all
necessary information. At WAL replay time, it's easy to re-execute
freezing because we don't need to re-compute the new infomask/Xmax
values but just take them from the WAL record.
While at it, restructure the coding to ensure all page changes occur in
a single critical section without much room for failures. The previous
coding wasn't using a critical section, without any explanation as to
why this was acceptable.
In replication scenarios using the 9.3 branch, standby servers must be
upgraded before their master, so that they are prepared to deal with the
new WAL record once the master is upgraded; failure to do so will cause
WAL replay to die with a PANIC message. Later upgrade of the standby
will allow the process to continue where it left off, so there's no
disruption of the data in the standby in any case. Standbys know how to
deal with the old WAL record, so it's okay to keep the master running
the old code for a while.
In master, the old freeze WAL record is gone, for cleanliness' sake;
there's no compatibility concern there.
Backpatch to 9.3, where the original bug was introduced and where the
previous fix was backpatched.
Álvaro Herrera and Andres Freund
The original performs too poorly; in some scenarios it shows way too
high while profiling. Try to make it a bit smarter to avoid excessive
cosst. In particular, make it have a maximum size, and have entries be
sorted in LRU order; once the max size is reached, evict the oldest
entry to avoid it from growing too large.
Per complaint from Andres Freund in connection with new tuple freezing
code.
WAL records of hint bit updates is useful to tools that want to examine
which pages have been modified. In particular, this is required to make
the pg_rewind tool safe (without checksums).
This can also be used to test how much extra WAL-logging would occur if
you enabled checksums, without actually enabling them (which you can't
currently do without re-initdb'ing).
Sawada Masahiko, docs by Samrat Revagade. Reviewed by Dilip Kumar, with
further changes by me.
Set min_recovery_apply_delay to force a delay in recovery apply for commit and
restore point WAL records. Other records are replayed immediately. Delay is
measured between WAL record time and local standby time.
Robert Haas, Fabrízio de Royes Mello and Simon Riggs
Detailed review by Mitsumasa Kondo
There's no real point in not doing this. It doesn't cost anything
in performance or space. So let's go wild.
Andres Freund, with substantial editing as to style by me.
When this reloption is set and wal_level=logical is configured,
we'll record the CIDs stamped by inserts, updates, and deletes to
the table just as we would for an actual catalog table. This will
allow logical decoding to use historical MVCC snapshots to access
such tables just as they access ordinary catalog tables.
Replication solutions built around the logical decoding machinery
will likely need to set this operation for their configuration
tables; it might also be needed by extensions which perform table
access in their output functions.
Andres Freund, reviewed by myself and others.
When wal_level=logical, we'll log columns from the old tuple as
configured by the REPLICA IDENTITY facility added in commit
07cacba983. This makes it possible
a properly-configured logical replication solution to correctly
follow table updates even if they change the chosen key columns,
or, with REPLICA IDENTITY FULL, even if the table has no key at
all. Note that updates which do not modify the replica identity
column won't log anything extra, making the choice of a good key
(i.e. one that will rarely be changed) important to performance
when wal_level=logical is configured.
Each insert, update, or delete to a catalog table will also log
the CMIN and/or CMAX values of stamped by the current transaction.
This is necessary because logical decoding will require access to
historical snapshots of the catalog in order to decode some data
types, and the CMIN/CMAX values that we may need in order to judge
row visibility may have been overwritten by the time we need them.
Andres Freund, reviewed in various versions by myself, Heikki
Linnakangas, KONDO Mitsumasa, and many others.
In 247c76a989, I added some code to do fine-grained checking of
MultiXact status of locking/updating transactions when traversing an
update chain. There was a thinko in that patch which would have the
traversing abort, that is return HeapTupleUpdated, when the other
transaction is a committed lock-only. In this case we should ignore it
and return success instead. Of course, in the case where there is a
committed update, HeapTupleUpdated is the correct return value.
A user-visible symptom of this bug is that in REPEATABLE READ and
SERIALIZABLE transaction isolation modes spurious serializability errors
can occur:
ERROR: could not serialize access due to concurrent update
In order for this to happen, there needs to be a tuple that's key-share-
locked and also updated, and the update must abort; a subsequent
transaction trying to acquire a new lock on that tuple would abort with
the above error. The reason is that the initial FOR KEY SHARE is seen
as committed by the new locking transaction, which triggers this bug.
(If the UPDATE commits, then the serialization error is correctly
reported.)
When running a query in READ COMMITTED mode, what happens is that the
locking is aborted by the HeapTupleUpdated return value, then
EvalPlanQual fetches the newest version of the tuple, which is then the
only version that gets locked. (The second time the tuple is checked
there is no misbehavior on the committed lock-only, because it's not
checked by the code that traverses update chains; so no bug.) Only the
newest version of the tuple is locked, not older ones, but this is
harmless.
The isolation test added by this commit illustrates the desired
behavior, including the proper serialization errors that get thrown.
Backpatch to 9.3.
Insertion to a non-leaf GIN page didn't make a full-page image of the page,
which is wrong. The code used to do it correctly, but was changed (commit
853d1c3103) because the redo-routine didn't
track incomplete splits correctly when the page was restored from a full
page image. Of course, that was not right way to fix it, the redo routine
should've been fixed instead. The redo-routine was surreptitiously fixed
in 2010 (commit 4016bdef8a), so all we need
to do now is revert the code that creates the record to its original form.
This doesn't change the format of the WAL record.
Backpatch to all supported versions.
When an external recovery command such as restore_command or
archive_cleanup_command fails, report the exit code properly,
distinguishing signals and normal exists, using the existing
wait_result_to_str() facility, instead of just reporting the return
value from system().
Reviewed-by: Peter Geoghegan <pg@heroku.com>
Both heap_freeze_tuple() and heap_tuple_needs_freeze() neglected to look
into a multixact to check the members against cutoff_xid. This means
that a very old Xid could survive hidden within a multi, possibly
outliving its CLOG storage. In the distant future, this would cause
clog lookup failures:
ERROR: could not access status of transaction 3883960912
DETAIL: Could not open file "pg_clog/0E78": No such file or directory.
This mostly was problematic when the updating transaction aborted, since
in that case the row wouldn't get pruned away earlier in vacuum and the
multixact could possibly survive for a long time. In many cases, data
that is inaccessible for this reason way can be brought back
heuristically.
As a second bug, heap_freeze_tuple() didn't properly handle multixacts
that need to be frozen according to cutoff_multi, but whose updater xid
is still alive. Instead of preserving the update Xid, it just set Xmax
invalid, which leads to both old and new tuple versions becoming
visible. This is pretty rare in practice, but a real threat
nonetheless. Existing corrupted rows, unfortunately, cannot be repaired
in an automated fashion.
Existing physical replicas might have already incorrectly frozen tuples
because of different behavior than in master, which might only become
apparent in the future once pg_multixact/ is truncated; it is
recommended that all clones be rebuilt after upgrading.
Following code analysis caused by bug report by J Smith in message
CADFUPgc5bmtv-yg9znxV-vcfkb+JPRqs7m2OesQXaM_4Z1JpdQ@mail.gmail.com
and privately by F-Secure.
Backpatch to 9.3, where freezing of MultiXactIds was introduced.
Analysis and patch by Andres Freund, with some tweaks by Álvaro.
It is dangerous to do so, because some code expects to be able to see what's
the true Xmax even if it is aborted (particularly while traversing HOT
chains). So don't do it, and instead rely on the callers to verify for
abortedness, if necessary.
Several race conditions and bugs fixed in the process. One isolation test
changes the expected output due to these.
This also reverts commit c235a6a589, which is no longer necessary.
Backpatch to 9.3, where this function was introduced.
Andres Freund
Commit 9dc842f08 of 8.2 era prevented MultiXact truncation during crash
recovery, because there was no guarantee that enough state had been
setup, and because it wasn't deemed to be a good idea to remove data
during crash recovery anyway. Since then, due to Hot-Standby, streaming
replication and PITR, the amount of time a cluster can spend doing crash
recovery has increased significantly, to the point that a cluster may
even never come out of it. This has made not truncating the content of
pg_multixact/ not defensible anymore.
To fix, take care to setup enough state for multixact truncation before
crash recovery starts (easy since checkpoints contain the required
information), and move the current end-of-recovery actions to a new
TrimMultiXact() function, analogous to TrimCLOG().
At some later point, this should probably done similarly to the way
clog.c is doing it, which is to just WAL log truncations, but we can't
do that for the back branches.
Back-patch to 9.0. 8.4 also has the problem, but since there's no hot
standby there, it's much less pressing. In 9.2 and earlier, this patch
is simpler than in newer branches, because multixact access during
recovery isn't required. Add appropriate checks to make sure that's not
happening.
Andres Freund
While autovacuum dutifully launched anti-multixact-wraparound vacuums
when the multixact "age" was reached, the vacuum code was not aware that
it needed to make them be full table vacuums. As the resulting
partial-table vacuums aren't capable of actually increasing relminmxid,
autovacuum continued to launch anti-wraparound vacuums that didn't have
the intended effect, until age of relfrozenxid caused the vacuum to
finally be a full table one via vacuum_freeze_table_age.
To fix, introduce logic for multixacts similar to that for plain
TransactionIds, using the same GUCs.
Backpatch to 9.3, where permanent MultiXactIds were introduced.
Andres Freund, some cleanup by Álvaro
Parts of the code used autovacuum_freeze_max_age to determine whether
anti-multixact-wraparound vacuums are necessary, while others used a
hardcoded 200000000 value. This leads to problems when
autovacuum_freeze_max_age is set to a non-default value. Use the latter
everywhere.
Backpatch to 9.3, where vacuuming of multixacts was introduced.
Andres Freund
Prevent handle_sig_alarm from losing control partway through due to a query
cancel (either an asynchronous SIGINT, or a cancel triggered by one of the
timeout handler functions). That would at least result in failure to
schedule any required future interrupt, and might result in actual
corruption of timeout.c's data structures, if the interrupt happened while
we were updating those.
We could still lose control if an asynchronous SIGINT arrives just as the
function is entered. This wouldn't break any data structures, but it would
have the same effect as if the SIGALRM interrupt had been silently lost:
we'd not fire any currently-due handlers, nor schedule any new interrupt.
To forestall that scenario, forcibly reschedule any pending timer interrupt
during AbortTransaction and AbortSubTransaction. We can avoid any extra
kernel call in most cases by not doing that until we've allowed
LockErrorCleanup to kill the DEADLOCK_TIMEOUT and LOCK_TIMEOUT events.
Another hazard is that some platforms (at least Linux and *BSD) block a
signal before calling its handler and then unblock it on return. When we
longjmp out of the handler, the unblock doesn't happen, and the signal is
left blocked indefinitely. Again, we can fix that by forcibly unblocking
signals during AbortTransaction and AbortSubTransaction.
These latter two problems do not manifest when the longjmp reaches
postgres.c, because the error recovery code there kills all pending timeout
events anyway, and it uses sigsetjmp(..., 1) so that the appropriate signal
mask is restored. So errors thrown outside any transaction should be OK
already, and cleaning up in AbortTransaction and AbortSubTransaction should
be enough to fix these issues. (We're assuming that any code that catches
a query cancel error and doesn't re-throw it will do at least a
subtransaction abort to clean up; but that was pretty much required already
by other subsystems.)
Lastly, ProcSleep should not clear the LOCK_TIMEOUT indicator flag when
disabling that event: if a lock timeout interrupt happened after the lock
was granted, the ensuing query cancel is still going to happen at the next
CHECK_FOR_INTERRUPTS, and we want to report it as a lock timeout not a user
cancel.
Per reports from Dan Wood.
Back-patch to 9.3 where the new timeout handling infrastructure was
introduced. We may at some point decide to back-patch the signal
unblocking changes further, but I'll desist from that until we hear
actual field complaints about it.
Although user-defined relations can't be directly created in
pg_catalog, it's possible for them to end up there, because you can
create them in some other schema and then use ALTER TABLE .. SET SCHEMA
to move them there. Previously, such relations couldn't afterwards
be manipulated, because IsSystemRelation()/IsSystemClass() rejected
all attempts to modify objects in the pg_catalog schema, regardless
of their origin. With this patch, they now reject only those
objects in pg_catalog which were created at initdb-time, allowing
most operations on user-created tables in pg_catalog to proceed
normally.
This patch also adds new functions IsCatalogRelation() and
IsCatalogClass(), which is similar to IsSystemRelation() and
IsSystemClass() but with a slightly narrower definition: only TOAST
tables of system catalogs are included, rather than *all* TOAST tables.
This is currently used only for making decisions about when
invalidation messages need to be sent, but upcoming logical decoding
patches will find other uses for this information.
Andres Freund, with some modifications by me.
In the GIN incomplete-splits patch, I used BlockIdDatas to store the block
number of left and right children, when inserting a downlink after a split
to an internal page posting list page. But gin_desc thought they were stored
as BlockNumbers.
Instead of simply checking the KEYS_UPDATED bit, we need to check
whether each lock held on the future version of the tuple conflicts with
the lock we're trying to acquire.
Per bug report #8434 by Tomonari Katsumata
Not doing so causes us to traverse an update chain that has been broken
by concurrent page pruning. All other code that traverses update chains
uses this check as one of the cases in which to stop iterating, so
replicate it here too. Failure to do so leads to erroneous CLOG,
subtrans or multixact lookups.
Per discussion following the bug report by J Smith in
CADFUPgc5bmtv-yg9znxV-vcfkb+JPRqs7m2OesQXaM_4Z1JpdQ@mail.gmail.com
as diagnosed by Andres Freund.
If a transaction updates/deletes a tuple just before aborting, and a
concurrent transaction tries to prune the page concurrently, the pruner
may see HeapTupleSatisfiesVacuum return HEAPTUPLE_DELETE_IN_PROGRESS,
but a later call to HeapTupleGetUpdateXid() return InvalidXid. This
would cause an assertion failure in development builds, but would be
otherwise Mostly Harmless.
Fix by checking whether the updater Xid is valid before trying to apply
it as page prune point.
Reported by Andres in 20131124000203.GA4403@alap2.anarazel.de
The reason for the fetch failure is that the tuple was removed because
it was dead; so the failure is innocuous and can be ignored. Moreover,
there's no need for further work and we can return success to the caller
immediately. EvalPlanQualFetch is doing something very similar to this
already.
Report and test case from Andres Freund in
20131124000203.GA4403@alap2.anarazel.de
Replace it with an approach similar to what GiST uses: when a page is split,
the left sibling is marked with a flag indicating that the parent hasn't been
updated yet. When the parent is updated, the flag is cleared. If an insertion
steps on a page with the flag set, it will finish split before proceeding
with the insertion.
The post-recovery cleanup mechanism was never totally reliable, as insertion
to the parent could fail e.g because of running out of memory or disk space,
leaving the tree in an inconsistent state.
This also divides the responsibility of WAL-logging more clearly between
the generic ginbtree.c code, and the parts specific to entry and posting
trees. There is now a common WAL record format for insertions and deletions,
which is written by ginbtree.c, followed by tree-specific payload, which is
returned by the placetopage- and split- callbacks.
Separate the insertion payload from the more static portions of GinBtree.
GinBtree now only contains information related to searching the tree, and
the information of what to insert is passed separately.
Add root block number to GinBtree, instead of passing it around all the
functions as argument.
Split off ginFinishSplit() from ginInsertValue(). ginFinishSplit is
responsible for finding the parent and inserting the downlink to it.
Change SET LOCAL/CONSTRAINTS/TRANSACTION behavior outside of a
transaction block from error (post-9.3) to warning. (Was nothing in <=
9.3.) Also change ABORT outside of a transaction block from notice to
warning.
These bugs can cause data loss on standbys started with hot_standby=on at
the moment they start to accept read only queries, by marking committed
transactions as uncommited. The likelihood of such corruptions is small
unless the primary has a high transaction rate.
5a031a5556 fixed bugs in HS's startup logic
by maintaining less state until at least STANDBY_SNAPSHOT_PENDING state
was reached, missing the fact that both clog and subtrans are written to
before that. This only failed to fail in common cases because the usage
of ExtendCLOG in procarray.c was superflous since clog extensions are
actually WAL logged.
f44eedc3f0f347a856eea8590730769125964597/I then tried to fix the missing
extensions of pg_subtrans due to the former commit's changes - which are
not WAL logged - by performing the extensions when switching to a state
> STANDBY_INITIALIZED and not performing xid assignments before that -
again missing the fact that ExtendCLOG is unneccessary - but screwed up
twice: Once because latestObservedXid wasn't updated anymore in that
state due to the earlier commit and once by having an off-by-one error in
the loop performing extensions. This means that whenever a
CLOG_XACTS_PER_PAGE (32768 with default settings) boundary was crossed
between the start of the checkpoint recovery started from and the first
xl_running_xact record old transactions commit bits in pg_clog could be
overwritten if they started and committed in that window.
Fix this mess by not performing ExtendCLOG() in HS at all anymore since
it's unneeded and evidently dangerous and by performing subtrans
extensions even before reaching STANDBY_SNAPSHOT_PENDING.
Analysis and patch by Andres Freund. Reported by Christophe Pettus.
Backpatch down to 9.0, like the previous commit that caused this.
RecoveryIsInProgress() can be called very frequently. During normal
operation, it just checks a backend-local variable and returns quickly,
but during hot standby, it checks a spinlock-protected shared variable.
Those spinlock acquisitions can become a point of contention on a busy
hot standby system.
Replace the spinlock acquisition with a memory barrier.
Per discussion with Andres Freund, Ants Aasma and Merlin Moncure.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
Split off the portion of ginInsertValue that inserts the tuple to current
level into a separate function, ginPlaceToPage. ginInsertValue's charter
is now to recurse up the tree to insert the downlink, when a page split is
required.
This is in preparation for a patch to change the way incomplete splits are
handled, which will need to do these operations separately. And IMHO makes
the code more readable anyway.
This creates a new gin-btree callback function for creating a downlink for
a page. Previously, ginxlog.c duplicated the logic used during normal
operation.
The root page is filled with as many items as fit, and the rest are inserted
using normal insertions. However, I fumbled the variable names, and the code
actually memcpy'd all the items on the page, overflowing the buffer. While
at it, rename the variable to make the distinction more clear.
Reported by Teodor Sigaev. This bug was introduced by my recent
refactorings, so no backpatching required.
If a page is deleted, and reused for something else, just as a search is
following a rightlink to it from its left sibling, the search would continue
scanning whatever the new contents of the page are. That could lead to
incorrect query results, or even something more curious if the page is
reused for a different kind of a page.
To fix, modify the search algorithm to lock the next page before releasing
the previous one, and refrain from deleting pages from the leftmost branch
of the tree.
Add a new Concurrency section to the README, explaining why this works.
There is a lot more one could say about concurrency in GIN, but that's for
another patch.
Backpatch to all supported versions.
Merge the isEnoughSpace and placeToPage functions in the b-tree interface
into one function that tries to put a tuple on page, and returns false if
it doesn't fit.
Move createPostingTree function to gindatapage.c, and change its contract
so that it can be passed more items than fit on the root page. It's in a
better position than the callers to know how many items fit.
Move ginMergeItemPointers out of gindatapage.c, into a separate file.
These changes make no difference now, but reduce the footprint of Alexander
Korotkov's upcoming patch to pack item pointers more tightly.
Historically, printtup() has assumed that it could prevent memory leakage
by pfree'ing the string result of each output function and manually
managing detoasting of toasted values. This amounts to assuming that
datatype output functions never leak any memory internally; an assumption
we've already decided to be bogus elsewhere, for example in COPY OUT.
range_out in particular is known to leak multiple kilobytes per call, as
noted in bug #8573 from Godfried Vanluffelen. While we could go in and fix
that leak, it wouldn't be very notationally convenient, and in any case
there have been and undoubtedly will again be other leaks in other output
functions. So what seems like the best solution is to run the output
functions in a temporary memory context that can be reset after each row,
as we're doing in COPY OUT. Some quick experimentation suggests this is
actually a tad faster than the retail pfree's anyway.
This patch fixes all the variants of printtup, except for debugtup()
which is used in standalone mode. It doesn't seem worth worrying
about query-lifespan leaks in standalone mode, and fixing that case
would be a bit tedious since debugtup() doesn't currently have any
startup or shutdown functions.
While at it, remove manual detoast management from several other
output-function call sites that had copied it from printtup(). This
doesn't make a lot of difference right now, but in view of recent
discussions about supporting "non-flattened" Datums, we're going to
want that code gone eventually anyway.
Back-patch to 9.2 where range_out was introduced. We might eventually
decide to back-patch this further, but in the absence of known major
leaks in older output functions, I'll refrain for now.
The original coding thought this case was impossible, but it can happen
if the bgwriter or checkpointer processes decide to write out an index
page while creation is still proceeding, leading to a bogus "unexpected
spgdoinsert() failure" error. Problem reported by Jonathan S. Katz.
Teodor Sigaev
The C and POSIX standards state that strncpy's behavior is undefined when
source and destination areas overlap. While it remains dubious whether any
implementations really misbehave when the pointers are exactly equal, some
platforms are now starting to force the issue by complaining when an
undefined call occurs. (In particular OS X 10.9 has been seen to dump core
here, though the exact set of circumstances needed to trigger that remain
elusive. Similar behavior can be expected to be optional on Linux and
other platforms in the near future.) So tweak the code to explicitly do
nothing when nothing need be done.
Back-patch to all active branches. In HEAD, this also lets us get rid of
an exception in valgrind.supp.
Per discussion of a report from Matthias Schmitt.