This requires adjusting the API for syscache callback functions: they now
get a hash value, not a TID, to identify the target tuple. Most of them
weren't paying any attention to that argument anyway, but plancache did
require a small amount of fixing.
Also, improve performance a trifle by avoiding sending duplicate inval
messages when a heap_update isn't changing the catcache lookup columns.
The TID isn't stable enough: we might queue an sinval event before a VACUUM
FULL, and then process it afterwards, when the target tuple no longer has
the same TID. So we must invalidate entries on the basis of hash value
only. The old coding can be shown to result in various bizarre,
hard-to-reproduce errors in the presence of concurrent VACUUM FULLs on
system catalogs, and could easily result in permanent catalog corruption,
up to and including complete loss of tables.
This commit is just a minimal fix that removes the unsafe comparison.
We should remove transmission of the tuple TID from sinval messages
altogether, and then arrange to suppress the extra message in the common
case of a heap_update that doesn't change the key hashvalue. But that's
going to be much more invasive, and will only produce a probably-marginal
performance gain, so it doesn't seem like material for a back-patch.
Back-patch to 9.0. Before that, VACUUM FULL refused to do any tuple moving
if it found any INSERT_IN_PROGRESS or DELETE_IN_PROGRESS tuples (and
CLUSTER would give up altogether), so there was no risk of moving a tuple
that might be the subject of an unsent sinval message.
We have to be sure that we have revalidated each nailed-in-cache relcache
entry before we try to use it to load data for some other relcache entry.
The introduction of "mapped relations" in 9.0 broke this, because although
we updated the state kept in relmapper.c early enough, we failed to
propagate that information into relcache entries soon enough; in
particular, we could try to fetch pg_class rows out of pg_class before
we'd updated its relcache entry's rd_node.relNode value from the map.
This bug accounts for Dave Gould's report of failures after "vacuum full
pg_class", and I believe that there is risk for other system catalogs
as well.
The core part of the fix is to copy relmapper data into the relcache
entries during "phase 1" in RelationCacheInvalidate(), before they'll be
used in "phase 2". To try to future-proof the code against other similar
bugs, I also rearranged the order in which nailed relations are visited
during phase 2: now it's pg_class first, then pg_class_oid_index, then
other nailed relations. This should ensure that RelationClearRelation can
apply RelationReloadIndexInfo to all nailed indexes without risking use
of not-yet-revalidated relcache entries.
Back-patch to 9.0 where the relation mapper was introduced.
This works around the problem that a catalog cache entry might contain a
toast pointer that we try to dereference just as a VACUUM FULL completes
on that catalog. We will see the sinval message on the cache entry when
we acquire lock on the toast table, but by that point we've already told
tuptoaster.c "here's the pointer to fetch", so it's difficult from a code
structural standpoint to update the pointer before we use it. Much less
painful to ensure that toast pointers are not invalidated in the first
place. We have to add a bit of code to deal with the case that a value
that previously wasn't toasted becomes so; but that should be a
seldom-exercised corner case, so the inefficiency shouldn't be significant.
Back-patch to 9.0. In prior versions, we didn't allow CLUSTER on system
catalogs, and VACUUM FULL didn't result in reassignment of toast OIDs, so
there was no problem.
The previous code tried to synchronize by unlinking the init file twice,
but that doesn't actually work: it leaves a window wherein a third process
could read the already-stale init file but miss the SI messages that would
tell it the data is stale. The result would be bizarre failures in catalog
accesses, typically "could not read block 0 in file ..." later during
startup.
Instead, hold RelCacheInitLock across both the unlink and the sending of
the SI messages. This is more straightforward, and might even be a bit
faster since only one unlink call is needed.
This has been wrong since it was put in (in 2002!), so back-patch to all
supported releases.
When streaming including WAL, the size estimate will always be incorrect,
since we don't know how much WAL is included. To make sure the output doesn't
look completely unreasonable, this patch increases the total size whenever we
go past the estimate, to make sure we never go above 100%.
This makes it clearer that the error message is perhaps not supposed
to be understood by users, and it also makes it somewhat clearer that
it was not accidentally omitted from translation.
Idea from Heikki Linnakangas, except that we don't mark "Reason code"
for translation at this point, because that would make the
implementation too cumbersome.
When updating or deleting a system catalog tuple, it's necessary to acquire
RowExclusiveLock on the catalog before looking up the tuple; otherwise a
concurrent VACUUM FULL on the catalog might move the tuple to a different
TID before we can apply the update. Coding patterns that find the tuple
via a table scan aren't at risk here, but when obtaining the tuple from a
catalog cache, correct ordering is important; and several routines in
foreigncmds.c got it wrong. Noted while running the regression tests in
parallel with VACUUM FULL of assorted system catalogs.
For consistency I moved all the heap_open calls to the starts of their
functions, including a couple for which there was no actual bug.
Back-patch to 8.4 where foreigncmds.c was added.
The statement start timestamp was not set before initiating the transaction
that is used to look up client authentication information in pg_authid.
In consequence, enable_sig_alarm computed a wrong value (far in the past)
for statement_fin_time. That didn't have any immediate effect, because the
timeout alarm was set without reference to statement_fin_time; but if we
subsequently blocked on a lock for a short time, CheckStatementTimeout
would consult the bogus value when we cancelled the lock timeout wait,
and then conclude we'd timed out, leading to immediate failure of the
connection attempt. Thus an innocent "vacuum full pg_authid" would cause
failures of concurrent connection attempts. Noted while testing other,
more serious consequences of vacuum full on system catalogs.
We should set the statement timestamp before StartTransactionCommand(),
so that the transaction start timestamp is also valid. I'm not sure if
there are any non-cosmetic effects of it not being valid, but the xact
timestamp is at least sent to the statistics machinery.
Back-patch to 9.0. Before that, the client authentication timeout was done
outside any transaction and did not depend on this state to be valid.
poll() is preferred over select() on platforms where both are available,
because it tends to be a bit faster and it doesn't have an arbitrary limit
on the range of FD numbers that can be accessed. The FD range limit does
not appear to be a risk factor for any 9.1 usages, so this doesn't need to
be back-patched, but we need to have it in place if we keep on expanding
the uses of WaitLatch.
Instead of displaying comments on an arbitrary subset of the object
types which support them, make \dd display comments on exactly those
object types which don't have their own backlash commands. We now
regard the display of comments as properly the job of the relevant
backslash command (though many of them do so only in verbose mode)
rather than something that \dd should be responsible for. However,
a handful of object types have no backlash command, so make \dd
give information about those.
Josh Kupershmidt
The latch infrastructure is now capable of detecting all cases where the
walsender loop needs to wake up, so there is no reason to have an arbitrary
timeout.
Also, modify the walsender loop logic to follow the standard pattern of
ResetLatch, test for work to do, WaitLatch. The previous coding was both
hard to follow and buggy: it would sometimes busy-loop despite having
nothing available to do, eg between receipt of a signal and the next time
it was caught up with new WAL, and it also had interesting choices like
deciding to update to WALSNDSTATE_STREAMING on the strength of information
known to be obsolete.
This is in hopes of learning more about what causes "pgstat wait timeout"
warnings in the buildfarm. This patch should probably be reverted once
we've learned what we can. As coded, it will result in regression test
"failures" at half the delay that the existing code does, so I expect
to see a few more than before.
In pursuit of this (and with the expectation that WaitLatch will be needed
in more places), convert the latch field that was already added to PGPROC
for sync rep into a generic latch that is activated for all PGPROC-owning
processes, and change many of the standard backend signal handlers to set
that latch when a signal happens. This will allow WaitLatch callers to be
wakened properly by these signals.
In passing, fix a whole bunch of signal handlers that had been hacked to do
things that might change errno, without adding the necessary save/restore
logic for errno. Also make some minor fixes in unix_latch.c, and clean
up bizarre and unsafe scheme for disowning the process's latch. Much of
this has to be back-patched into 9.1.
Peter Geoghegan, with additional work by Tom
streamed backup, throw an error and refuse to start up. The restore has not
finished correctly in that case and the data directory is possibly corrupt.
We already errored out in case of archive recovery, but could not during
crash recovery because we couldn't distinguish between the case that
pg_start_backup() was called and the database then crashed (must not error,
data is OK), and the case that we're restoring from a backup and not all
the needed WAL was replayed (data can be corrupt).
To distinguish those cases, add a line to backup_label to indicate
whether the backup was taken with pg_start/stop_backup(), or by streaming
(ie. pg_basebackup).
This requires re-initdb, because of a new field added to the control file.
The original definition had the problem that timeouts exceeding about 2100
seconds couldn't be specified on 32-bit machines. Milliseconds seem like
sufficient resolution, and finer grain than that would be fantasy anyway
on many platforms.
Back-patch to 9.1 so that this aspect of the latch API won't change between
9.1 and later releases.
Peter Geoghegan
Improve the documentation around weak-memory-ordering risks, and do a pass
of general editorialization on the comments in the latch code. Make the
Windows latch code more like the Unix latch code where feasible; in
particular provide the same Assert checks in both implementations.
Fix poorly-placed WaitLatch call in syncrep.c.
This patch resolves, for the moment, concerns around weak-memory-ordering
bugs in latch-related code: we have documented the restrictions and checked
that existing calls meet them. In 9.2 I hope that we will install suitable
memory barrier instructions in SetLatch/ResetLatch, so that their callers
don't need to be quite so careful.
Such a construction is useless since the lower PlaceHolderVar is already
nullable; no need to make it more so. Noted while pursuing bug #6154.
This is just a minor planner efficiency improvement, since the final plan
will come out the same anyway after PHVs are flattened. So not worth the
risk of back-patching.
Don't try to allocate the default value for a string relopt in the same
palloc chunk as the relopt_string struct. That didn't work too well if you
added a built-in string relopt in the stringRelOpts array, as it's not
possible to have an initializer for a variable length struct in C. This
makes the code slightly simpler too.
While we're at it, move the call to validator function in
add_string_reloption to before the allocation, so that if someone does pass
a bogus default value, we don't leak memory.
A PlaceHolderVar's expression might contain another, lower-level
PlaceHolderVar. If the outer PlaceHolderVar is used, the inner one
certainly will be also, and so we have to make sure that both of them get
into the placeholder_list with correct ph_may_need values during the
initial pre-scan of the query (before deconstruct_jointree starts).
We did this correctly for PlaceHolderVars appearing in the query quals,
but overlooked the issue for those appearing in the top-level targetlist;
with the result that nested placeholders referenced only in the targetlist
did not work correctly, as illustrated in bug #6154.
While at it, add some error checking to find_placeholder_info to ensure
that we don't try to create new placeholders after it's too late to do so;
they have to all be created before deconstruct_jointree starts.
Back-patch to 8.4 where the PlaceHolderVar mechanism was introduced.
The relevant backslash commands already exist, so we're just adding an
additional column. With this commit, all objects that have psql backslash
commands and accept comments should now display those comments at least
in verbose mode.
Josh Kupershmidt, with doc additions by me.
\dc and \dD now accept a "+" option, which will cause the comments to
be displayed. Along the way, correct a few oversights in the previous
commit in this area, 3b17efdfdd - namely,
(1) when \dL+ is used, make description still be the last column, for
consistency with what we've done elsewhere; and (2) document the
difference between \dC and \dC+.
Josh Kupershmidt, with a couple of doc changes by me.
This lie has been harmless until now, but has been exposed by the
change to include postgres.h before the python headers, which
in some versions include inttypes.h if HAVE_INTTYPES_H is set.
Somebody thought it'd be cute to invent a set of Node tag numbers that were
defined independently of, and indeed conflicting with, the main tag-number
list. While this accidentally failed to fail so far, it would certainly
lead to trouble as soon as anyone wanted to, say, apply copyObject to these
node types. Clang was already complaining about the use of makeNode on
these tags, and I think quite rightly so. Fix by pushing these node
definitions into the mainstream, including putting replnodes.h where it
belongs.
The previous limit of 1024 was set on the assumption that all modern syslog
implementations have line length limits of 2KB or so. However, this is
false, as at least Solaris and sysklogd truncate at only 1KB. 900 seems
to leave enough room for the max likely length of the tacked-on prefixes,
so let's go with that.
As with the previous change, it doesn't seem wise to back-patch this into
already-released branches; but it should be OK to sneak it into 9.1.
Noah Misch
The previous test for status < 0 test is in fact testing nothing if the
compiler considers an enum to be an unsigned data type. clang doesn't
like tautologies, so do this instead.
Report by Peter Geoghegan, fix as suggested by Tom Lane.
To avoid having the python headers hijack various definitions,
we now include them after all the system headers we want, having
first undefined some of the things they want to define. After that's
done we restore the things they scribbled on that matter, namely our
snprintf and vsnprintf macros, if we're using them.
Instead of entering them on transaction startup, we materialize them
only when someone wants to wait, which will occur only during CREATE
INDEX CONCURRENTLY. In Hot Standby mode, the startup process must also
be able to probe for conflicting VXID locks, but the lock need never be
fully materialized, because the startup process does not use the normal
lock wait mechanism. Since most VXID locks never need to touch the
lock manager partition locks, this can significantly reduce blocking
contention on read-heavy workloads.
Patch by me. Review by Jeff Davis.
The output of \dL (list languages) is fairly narrow, so we just always
display the comment. \dC (list casts) can get fairly wide, so we only
display comments if the new \dC+ option is specified.
Josh Kupershmidt
glibc renders random() thread-safe by wrapping a futex lock around it;
testing reveals that this limits the performance of pgbench on machines
with many CPU cores. Rather than switching to random_r(), which is
only available on GNU systems and crashes unless you use undocumented
alchemy to initialize the random state properly, switch to our built-in
implementation of erand48(), which is both thread-safe and concurrent.
Since the list of reasons not to use the operating system's erand48()
is getting rather long, rename ours to pg_erand48() (and similarly
for our implementations of lrand48() and srand48()) and just always
use those. We were already doing this on Cygwin anyway, and the
glibc implementation is not quite thread-safe, so pgbench wouldn't
be able to use that either.
Per discussion with Tom Lane.
This kluge was inserted in a spot apparently chosen at random: the lock
manager's state is not yet fully set up for the wait, and in particular
LockWaitCancel hasn't been armed by setting lockAwaited, so the ProcLock
will not get cleaned up if the ereport is thrown. This seems to not cause
any observable problem in trivial test cases, because LockReleaseAll will
silently clean up the debris; but I was able to cause failures with tests
involving subtransactions.
Fixes breakage induced by commit c85c941470.
Back-patch to all affected branches.
It was initialized in the wrong place and to the wrong value. With bad
luck this could result in incorrect query-cancellation failures in hot
standby sessions, should a HS backend be holding pin on buffer number 1
while trying to acquire a lock.
Testing shows that the overhead of acquiring and releasing
SInvalReadLock and msgNumLock on high-core count boxes can waste a lot
of CPU time and hurt performance. This patch adds a per-backend flag
that allows us to skip all that locking in most cases. Further
testing shows that this improves performance even when sinval traffic
is very high.
Patch by me. Review and testing by Noah Misch.
Heikki Linnakangas