We allegedly allocated all remaining memory for the read buffers of the
sort tapes, but we allocated the merge heap only after that. That means
that the allocation of the merge heap was guaranteed to go over the memory
limit. Fix by allocating the merge heap first. This makes little difference
in practice, because the merge heap is tiny, but let's tidy.
While we're at it, add a safeguard for the case that we are already over
the limit when allocating the read buffers. That shouldn't happen, but
better safe than sorry.
The memory accounting error was reported off-list by Peter Geoghegan.
Have tuplesort_gettupleslot() copy the contents of its current table slot
as needed. This is based on an approach taken by tuplestore_gettupleslot().
In the future, tuplesort_gettupleslot() may also be taught to avoid copying
the tuple where caller can determine that that is safe (the
tuplestore_gettupleslot() interface already offers this option to callers).
Patch by Peter Geoghegan. Fixes bug #14344, reported by Regina Obe.
Report: <20160929035538.20224.39628@wrigleys.postgresql.org>
Backpatch-through: 9.6
Pass the buffer size as argument to LogicalTapeRewindForRead, rather than
setting it earlier with the separate LogicTapeAssignReadBufferSize call.
This way, the buffer size is set closer to where it's actually used, which
makes the code easier to understand.
This makes the calculation for how much memory to use for the buffers less
precise. We now use the same amount of memory for every tape, rounded down
to the nearest BLCKSZ boundary, instead of using one more block for some
tapes, to get the total up to exact amount of memory available. That should
be OK, merging isn't too sensitive to the exact amount of memory used.
Reviewed by Peter Geoghegan
Discussion: <0f607c4b-df23-353e-bf56-c0389d28495f@iki.fi>
LogicalTapeRewind() should not allocate large read buffer, if the tape
is completely empty. The calling code relies on that, for its
calculation of how much memory to allocate for the read buffers. That
lead to massive overallocation of memory, if maxTapes was high, but
only a few tapes were actually used.
Reported by Tomas Vondra
Discussion: <7303da46-daf7-9c68-3cc1-9f83235cf37e@2ndquadrant.com>
Don't pre-read tuples into SortTuple slots during merge. Instead, use the
memory for larger read buffers in logtape.c. We're doing the same number
of READTUP() calls either way, but managing the pre-read SortTuple slots
is much more complicated. Also, the on-tape representation is more compact
than SortTuples, so we can fit more pre-read tuples into the same amount
of memory this way. And we have better cache-locality, when we use just a
small number of SortTuple slots.
Now that we only hold one tuple from each tape in the SortTuple slots, we
can greatly simplify the "batch memory" management. We now maintain a
small set of fixed-sized slots, to hold the tuples, and fall back to
palloc() for larger tuples. We use this method during all merge phases,
not just the final merge, and also when randomAccess is requested, and
also in the TSS_SORTEDONTAPE case. In other words, it's used whenever we
do an external sort.
Reviewed by Peter Geoghegan and Claudio Freire.
Discussion: <CAM3SWZTpaORV=yQGVCG8Q4axcZ3MvF-05xe39ZvORdU9JcD6hQ@mail.gmail.com>
In external sort's merge phase, we maintain a binary heap holding the next
tuple from each input tape. On each step, the topmost tuple is returned,
and replaced with the next tuple from the same tape. We were doing the
replacement by deleting the top node in one operation, and inserting the
next tuple after that. However, you can do a "replace-top" operation more
efficiently, in one "sift-up". A deletion will always walk the heap from
top to bottom, but in a replacement, we can stop as soon as we find the
right place for the new tuple. This is particularly helpful, if the tapes
are not in completely random order, so that the next tuple from a tape is
likely to land near the top of the heap.
Peter Geoghegan, reviewed by Claudio Freire, with some editing by me.
Discussion: <CAM3SWZRhBhiknTF_=NjDSnNZ11hx=U_SEYwbc5vd=x7M4mMiCw@mail.gmail.com>
Negative availMemLessRefund would be problematic. It's not entirely
clear whether the case can be hit in the code as it stands, but this
seems like good future-proofing in any case. While we're at it,
insist that the value be not merely positive but not tiny, so as to
avoid doing a lot of repalloc work for little gain.
Peter Geoghegan
Discussion: <CAM3SWZRVkuUB68DbAkgw=532gW0f+fofKueAMsY7hVYi68MuYQ@mail.gmail.com>
I found that half a dozen (nearly 5%) of our AllocSetContextCreate calls
had typos in the context-sizing parameters. While none of these led to
especially significant problems, they did create minor inefficiencies,
and it's now clear that expecting people to copy-and-paste those calls
accurately is not a great idea. Let's reduce the risk of future errors
by introducing single macros that encapsulate the common use-cases.
Three such macros are enough to cover all but two special-purpose contexts;
those two calls can be left as-is, I think.
While this patch doesn't in itself improve matters for third-party
extensions, it doesn't break anything for them either, and they can
gradually adopt the simplified notation over time.
In passing, change TopMemoryContext to use the default allocation
parameters. Formerly it could only be extended 8K at a time. That was
probably reasonable when this code was written; but nowadays we create
many more contexts than we did then, so that it's not unusual to have a
couple hundred K in TopMemoryContext, even without considering various
dubious code that sticks other things there. There seems no good reason
not to let it use growing blocks like most other contexts.
Back-patch to 9.6, mostly because that's still close enough to HEAD that
it's easy to do so, and keeping the branches in sync can be expected to
avoid some future back-patching pain. The bugs fixed by these changes
don't seem to be significant enough to justify fixing them further back.
Discussion: <21072.1472321324@sss.pgh.pa.us>
Due to an error in the abbreviated key abort logic, the most recently
processed SortTuple could be incorrectly marked NULL, resulting in an
incorrect final sort order.
In the worst case, this could result in a corrupt btree index, which
would need to be rebuild using REINDEX. However, abbrevation doesn't
abort very often, not all data types use it, and only one tuple would
end up in the wrong place, so the practical impact of this mistake may
be somewhat limited.
Report and patch by Peter Geoghegan.
Otherwise, when we abandon incremental memory accounting and use
batch allocation for the final merge pass, we might crash. This
has been broken since 0011c0091e.
Peter Geoghegan, tested by Noah Misch
Now indexes (but only B-tree for now) can contain "extra" column(s) which
doesn't participate in index structure, they are just stored in leaf
tuples. It allows to use index only scan by using single index instead
of two or more indexes.
Author: Anastasia Lubennikova with minor editorializing by me
Reviewers: David Rowley, Peter Geoghegan, Jeff Janes
We still use replacement selection for the first run of the sort only
and only when the number of tuples is relatively small. Otherwise,
the first run, and subsequent runs in all cases, are produced using
quicksort. This tends to be faster except perhaps for very small
amounts of working memory.
Peter Geoghegan, reviewed by Tomas Vondra, Jeff Janes, Mithun Cy,
Greg Stark, and me.
I'm committing these changes separately so that it's clear what is
Peter's original work versus what I changed. This is a followup to
commit 0011c0091e, and these changes
are all by me.
Introduce a new memory context which stores tuple data, and reset it
at the end of each merge pass; this helps avoid memory fragmentation
and, consequently, overallocation. Also, for the final merge patch,
eliminate memory context chunk header overhead entirely by allocating
all of the memory used for buffering tuples during the merge in a
single chunk. Since this modestly increases the number of tuples we
can store, grow the memtuples array a bit so that we're less likely to
run short of slots there.
Peter Geoghegan. Review and testing of patches in this series by
Jeff Janes, Greg Stark, Mithun Cy, and me.
When processing ordered aggregates following a sort that could make use
of the abbreviated key optimization, only call the equality operator to
compare successive pairs of tuples when their abbreviated keys were not
equal.
Peter Geoghegan, reviewd by Andreas Karlsson and by me.
This patch reduces pg_am to just two columns, a name and a handler
function. All the data formerly obtained from pg_am is now provided
in a C struct returned by the handler function. This is similar to
the designs we've adopted for FDWs and tablesample methods. There
are multiple advantages. For one, the index AM's support functions
are now simple C functions, making them faster to call and much less
error-prone, since the C compiler can now check function signatures.
For another, this will make it far more practical to define index access
methods in installable extensions.
A disadvantage is that SQL-level code can no longer see attributes
of index AMs; in particular, some of the crosschecks in the opr_sanity
regression test are no longer possible from SQL. We've addressed that
by adding a facility for the index AM to perform such checks instead.
(Much more could be done in that line, but for now we're content if the
amvalidate functions more or less replace what opr_sanity used to do.)
We might also want to expose some sort of reporting functionality, but
this patch doesn't do that.
Alexander Korotkov, reviewed by Petr Jelínek, and rather heavily
editorialized on by me.
So far we have worked around the fact that some very old compilers do
not support 'inline' functions by only using inline functions
conditionally (or not at all). Since such compilers are very rare by
now, we have decided to rely on inline functions from 9.6 onwards.
To avoid breaking these old compilers inline is defined away when not
supported. That'll cause "function x defined but not used" type of
warnings, but since nobody develops on such compilers anymore that's
ok.
This change in policy will allow us to more easily employ inline
functions.
I chose to remove code previously conditional on PG_USE_INLINE as it
seemed confusing to have code dependent on a define that's always
defined.
Blacklisting of compilers, like in c53f73879f, now has to be done
differently. A platform template can define PG_FORCE_DISABLE_INLINE to
force inline to be defined empty.
Discussion: 20150701161447.GB30708@awork2.anarazel.de
The tuplesort/tuplestore memory management logic assumed that the chunk
allocation overhead for its memtuples array could not increase when
increasing the array size. This is and always was true for tuplesort,
but we (I, I think) blindly copied that logic into tuplestore.c without
noticing that the assumption failed to hold for the much smaller array
elements used by tuplestore. Given rather small work_mem, this could
result in an improper complaint about "unexpected out-of-memory situation",
as reported by Brent DeSpain in bug #13530.
The easiest way to fix this is just to increase tuplestore's initial
array size so that the assumption holds. Rather than relying on magic
constants, though, let's export a #define from aset.c that represents
the safe allocation threshold, and make tuplestore's calculation depend
on that.
Do the same in tuplesort.c to keep the logic looking parallel, even though
tuplesort.c isn't actually at risk at present. This will keep us from
breaking it if we ever muck with the allocation parameters in aset.c.
Back-patch to all supported versions. The error message doesn't occur
pre-9.3, not so much because the problem can't happen as because the
pre-9.3 tuplestore code neglected to check for it. (The chance of
trouble is a great deal larger as of 9.3, though, due to changes in the
array-size-increasing strategy.) However, allowing LACKMEM() to become
true unexpectedly could still result in less-than-desirable behavior,
so let's patch it all the way back.
It's standard for quicksort implementations, after having partitioned the
input into two subgroups, to recurse to process the smaller partition and
then handle the larger partition by iterating. This method guarantees
that no more than log2(N) levels of recursion can be needed. However,
Bentley and McIlroy argued that checking to see which partition is smaller
isn't worth the cycles, and so their code doesn't do that but just always
recurses on the left partition. In most cases that's fine; but with
worst-case input we might need O(N) levels of recursion, and that means
that qsort could be driven to stack overflow. Such an overflow seems to
be the only explanation for today's report from Yiqing Jin of a SIGSEGV
in med3_tuple while creating an index of a couple billion entries with a
very large maintenance_work_mem setting. Therefore, let's spend the few
additional cycles and lines of code needed to choose the smaller partition
for recursion.
Also, fix up the qsort code so that it properly uses size_t not int for
some intermediate values representing numbers of items. This would only
be a live risk when sorting more than INT_MAX bytes (in qsort/qsort_arg)
or tuples (in qsort_tuple), which I believe would never happen with any
caller in the current core code --- but perhaps it could happen with
call sites in third-party modules? In any case, this is trouble waiting
to happen, and the corrected code is probably if anything shorter and
faster than before, since it removes sign-extension steps that had to
happen when converting between int and size_t.
In passing, move a couple of CHECK_FOR_INTERRUPTS() calls so that it's
not necessary to preserve the value of "r" across them, and prettify
the output of gen_qsort_tuple.pl a little.
Back-patch to all supported branches. The odds of hitting this issue
are probably higher in 9.4 and up than before, due to the new ability
to allocate sort workspaces exceeding 1GB, but there's no good reason
to believe that it's impossible to crash older branches this way.
Commit 5cefbf5a6c introduced an
assumption that this field would always be non-NULL when doing a merge
pass, but that's not true. Without this fix, you can crash the server
by building a hash index that is sufficiently large relative to
maintenance_work_mem, or by triggering a large datum sort.
Commit 5ea86e6e65 changed the comments
for that field to say that it would be set in all cases except for the
hash index case, but that wasn't (and still isn't) true.
The datum-sort failure was spotted by Tomas Vondra; initial analysis
of that failure was by Peter Geoghegan. The remaining issues were
spotted by me during review of the surrounding code, and the patch is
all my fault.
While building error messages to return to the user,
BuildIndexValueDescription, ExecBuildSlotValueDescription and
ri_ReportViolation would happily include the entire key or entire row in
the result returned to the user, even if the user didn't have access to
view all of the columns being included.
Instead, include only those columns which the user is providing or which
the user has select rights on. If the user does not have any rights
to view the table or any of the columns involved then no detail is
provided and a NULL value is returned from BuildIndexValueDescription
and ExecBuildSlotValueDescription. Note that, for key cases, the user
must have access to all of the columns for the key to be shown; a
partial key will not be returned.
Further, in master only, do not return any data for cases where row
security is enabled on the relation and row security should be applied
for the user. This required a bit of refactoring and moving of things
around related to RLS- note the addition of utils/misc/rls.c.
Back-patch all the way, as column-level privileges are now in all
supported versions.
This has been assigned CVE-2014-8161, but since the issue and the patch
have already been publicized on pgsql-hackers, there's no point in trying
to hide this commit.
When we write tuples out to disk and read them back in, the abbreviated
keys become non-abbreviated, because the readtup routines don't know
anything about abbreviation. But without this fix, the rest of the
code still thinks the abbreviation-aware compartor should be used,
so chaos ensues.
Report by Andrew Gierth; patch by Peter Geoghegan.
This commit extends the SortSupport infrastructure to allow operator
classes the option to provide abbreviated representations of Datums;
in the case of text, we abbreviate by taking the first few characters
of the strxfrm() blob. If the abbreviated comparison is insufficent
to resolve the comparison, we fall back on the normal comparator.
This can be much faster than the old way of doing sorting if the
first few bytes of the string are usually sufficient to resolve the
comparison.
There is the potential for a performance regression if all of the
strings to be sorted are identical for the first 8+ characters and
differ only in later positions; therefore, the SortSupport machinery
now provides an infrastructure to abort the use of abbreviation if
it appears that abbreviation is producing comparatively few distinct
keys. HyperLogLog, a streaming cardinality estimator, is included in
this commit and used to make that determination for text.
Peter Geoghegan, reviewed by me.
xlog.c is huge, this makes it a little bit smaller, which is nice. Functions
related to putting together the WAL record are in xloginsert.c, and the
lower level stuff for managing WAL buffers and such are in xlog.c.
Also move the definition of XLogRecord to a separate header file. This
causes churn in the #includes of all the files that write WAL records, and
redo routines, but it avoids pulling in xlog.h into most places.
Reviewed by Michael Paquier, Alvaro Herrera, Andres Freund and Amit Kapila.
This could be useful for datatypes like text, where we might want
to optimize for some collations but not others. However, this patch
doesn't introduce any new sortsupport functions that work this way;
it merely revises the code so that future patches may do so.
Patch by me. Review by Peter Geoghegan.
The previous code, perhaps out of concern for avoid memory leaks, formed
the tuple in one memory context and then copied it to another memory
context. However, this doesn't appear to be necessary, since
index_form_tuple and the functions it calls take precautions against
leaking memory. In my testing, building the tuple directly inside the
sort context shaves several percent off the index build time.
Rearrange things so we do that.
Patch by me. Review by Amit Kapila, Tom Lane, Andres Freund.
We should report the errno when we get a failure from functions like
BufFileWrite. "ERROR: write failed" is unreasonably taciturn for a
case that's well within the realm of possibility; I've seen it a
couple times in the buildfarm recently, in situations that were
probably out-of-disk-space, but it'd be good to see the errno
to confirm it.
I think this code was originally written without assuming that
the buffile.c functions would return useful errno; but most other
callers *are* assuming that, and a quick look at the buffile code
gives no reason to suppose otherwise.
Also, a couple of the old messages were phrased on the assumption
that a short read might indicate a logic bug in tuplestore itself;
but that code's pretty well tested by now, so a filesystem-level
problem seems much more likely.
Repositioning the tuplestore seek pointer in window_gettupleslot() turns
out to be a very significant expense when the window frame is sizable and
the frame end can move. To fix, introduce a tuplestore function for
skipping an arbitrary number of tuples in one call, parallel to the one we
introduced for tuplesort objects in commit 8d65da1f. This reduces the cost
of window_gettupleslot() to O(1) if the tuplestore has not spilled to disk.
As in the previous commit, I didn't try to do any real optimization of
tuplestore_skiptuples for the case where the tuplestore has spilled to
disk. There is probably no practical way to get the cost to less than O(N)
anyway, but perhaps someone can think of something later.
Also fix PersistHoldablePortal() to make use of this API now that we have
it.
Based on a suggestion by Dean Rasheed, though this turns out not to look
much like his patch.
This patch introduces generic support for ordered-set and hypothetical-set
aggregate functions, as well as implementations of the instances defined in
SQL:2008 (percentile_cont(), percentile_disc(), rank(), dense_rank(),
percent_rank(), cume_dist()). We also added mode() though it is not in the
spec, as well as versions of percentile_cont() and percentile_disc() that
can compute multiple percentile values in one pass over the data.
Unlike the original submission, this patch puts full control of the sorting
process in the hands of the aggregate's support functions. To allow the
support functions to find out how they're supposed to sort, a new API
function AggGetAggref() is added to nodeAgg.c. This allows retrieval of
the aggregate call's Aggref node, which may have other uses beyond the
immediate need. There is also support for ordered-set aggregates to
install cleanup callback functions, so that they can be sure that
infrastructure such as tuplesort objects gets cleaned up.
In passing, make some fixes in the recently-added support for variadic
aggregates, and make some editorial adjustments in the recent FILTER
additions for aggregates. Also, simplify use of IsBinaryCoercible() by
allowing it to succeed whenever the target type is ANY or ANYELEMENT.
It was inconsistent that it dealt with other polymorphic target types
but not these.
Atri Sharma and Andrew Gierth; reviewed by Pavel Stehule and Vik Fearing,
and rather heavily editorialized upon by Tom Lane
In tuplesort.c:inittapes(), we calculate tapeSpace by first figuring
out how many 'tapes' we can use (maxTapes) and then multiplying the
result by the tape buffer overhead for each. Unfortunately, when
we are on a system with an 8-byte long, we allow work_mem to be
larger than 2GB and that allows maxTapes to be large enough that the
32bit arithmetic can overflow when multiplied against the buffer
overhead.
When this overflow happens, we end up adding the overflow to the
amount of space available, causing the amount of memory allocated to
be larger than work_mem.
Note that to reach this point, you have to set work mem to at least
24GB and be sorting a set which is at least that size. Given that a
user who can set work_mem to 24GB could also set it even higher, if
they were looking to run the system out of memory, this isn't
considered a security issue.
This overflow risk was found by the Coverity scanner.
Back-patch to all supported branches, as this issue has existed
since before 8.4.
Commit 263865a489 switched tuplesort.c and
tuplestore.c variables representing memory usage from type "long" to
type "Size". This was unnecessary; I thought doing so avoided overflow
scenarios on 64-bit Windows, but guc.c already limited work_mem so as to
prevent the overflow. It was also incomplete, not touching the logic
that assumed a signed data type. Change the affected variables to
"int64". This is perfect for 64-bit platforms, and it reduces the need
to contemplate platform-specific overflow scenarios. It also puts us
close to being able to support work_mem over 2 GiB on 64-bit Windows.
Per report from Andres Freund.
The MaxAllocSize guard is convenient for most callers, because it
reduces the need for careful attention to overflow, data type selection,
and the SET_VARSIZE() limit. A handful of callers are happy to navigate
those hazards in exchange for the ability to allocate a larger chunk.
Introduce MemoryContextAllocHuge() and repalloc_huge(). Use this in
tuplesort.c and tuplestore.c, enabling internal sorts of up to INT_MAX
tuples, a factor-of-48 increase. In particular, B-tree index builds can
now benefit from much-larger maintenance_work_mem settings.
Reviewed by Stephen Frost, Simon Riggs and Jeff Janes.
This patch addresses the problem that applications currently have to
extract object names from possibly-localized textual error messages,
if they want to know for example which index caused a UNIQUE_VIOLATION
failure. It adds new error message fields to the wire protocol, which
can carry the name of a table, table column, data type, or constraint
associated with the error. (Since the protocol spec has always instructed
clients to ignore unrecognized field types, this should not create any
compatibility problem.)
Support for providing these new fields has been added to just a limited set
of error reports (mainly, those in the "integrity constraint violation"
SQLSTATE class), but we will doubtless add them to more calls in future.
Pavel Stehule, reviewed and extensively revised by Peter Geoghegan, with
additional hacking by Tom Lane.
The code originally just doubled the size of the tuple-pointer array so
long as that would fit in allowedMem. This could result in failing to use
as much as half of allowedMem, if (as is typical) the last doubling attempt
didn't quite fit. Worse, we might double the array size but be unable to
use most of the added slots, because there was no room left within the
allowedMem limit for tuples the slots should point to. To fix, double only
so long as we've used less than half of allowedMem in total. Then do one
more array enlargement, but scale it based on total memory consumption so
far. This will work nicely as long as the average tuple size is reasonably
stable, and in any case should be better than the old method.
This change will result in large sort operations consuming a larger
fraction of work_mem than they typically did in the past. The release
notes should mention that users may want to revisit their work_mem
settings, if they'd tuned those settings based on the old behavior of
sorting.
Jeff Janes, reviewed by Peter Geoghegan and Robert Haas
We already had those, but they forced modules to spell out the function
bodies twice. Eliminate some duplicates we had already grown.
Extracted from a somewhat larger patch from Andres Freund.
This reduces unnecessary exposure of other headers through htup.h, which
is very widely included by many files.
I have chosen to move the function prototypes to the new file as well,
because that means htup.h no longer needs to include tupdesc.h. In
itself this doesn't have much effect in indirect inclusion of tupdesc.h
throughout the tree, because it's also required by execnodes.h; but it's
something to explore in the future, and it seemed best to do the htup.h
change now while I'm busy with it.
Commit 337b6f5ecf contained the entirely
fanciful assumption that it had made comparetup_datum unreachable.
Reported and patched by Takashi Yamamoto.
Fix up some not terribly accurate/useful comments from that commit, too.
I broke this in commit 337b6f5ecf, which
among other things arranged for quicksorts to CHECK_FOR_INTERRUPTS()
slightly less frequently. Sadly, it also arranged for heapsorts to
CHECK_FOR_INTERRUPTS() much less frequently. Repair.
Per recent work by Peter Geoghegan, it's significantly faster to
tuplesort on a single sortkey if ApplySortComparator is inlined into
quicksort rather reached via a function pointer. It's also faster
in general to have a version of quicksort which is specialized for
sorting SortTuple objects rather than objects of arbitrary size and
type. This requires a couple of additional copies of the quicksort
logic, which in this patch are generate using a Perl script. There
might be some benefit in adding further specializations here too,
but thus far it's not clear that those gains are worth their weight
in code footprint.
Our own qsort_arg() implementation doesn't have the defect previously
observed to affect only QNX 4, so it seems sufficiently to assert that
it isn't broken rather than retesting. Also, update a few comments to
clarify why it's valuable to retain a tie-break rule based on CTID
during index builds.
Peter Geoghegan, with slight tweaks by me.
This patch creates an API whereby a btree index opclass can optionally
provide non-SQL-callable support functions for sorting. In the initial
patch, we only use this to provide a directly-callable comparator function,
which can be invoked with a bit less overhead than the traditional
SQL-callable comparator. While that should be of value in itself, the real
reason for doing this is to provide a datatype-extensible framework for
more aggressive optimizations, as in Peter Geoghegan's recent work.
Robert Haas and Tom Lane
This oversight could result in a tuplestore using much more than the
intended amount of memory. It would only happen in a code path that loaded
a tuplestore via tuplestore_putvalues(), and many of those won't emit huge
amounts of data; but cases such as holdable cursors and plpgsql's RETURN
NEXT command could have the problem. The fix ensures that the tuplestore
will switch to write-to-disk mode when it overruns work_mem.
The potential overrun was finite, because we would still count the space
used by the tuple pointer array, so the tuplestore code would eventually
flip into write-to-disk mode anyway. When storing wide tuples we would
go far past the expected work_mem usage before that happened; but this
may account for the lack of prior reports.
Back-patch to 8.4, where tuplestore_putvalues was introduced.
Per bug #6061 from Yann Delorme.
Since collation is effectively an argument, not a property of the function,
FmgrInfo is really the wrong place for it; and this becomes critical in
cases where a cached FmgrInfo is used for varying purposes that might need
different collation settings. Fix by passing it in FunctionCallInfoData
instead. In particular this allows a clean fix for bug #5970 (record_cmp
not working). This requires touching a bit more code than the original
method, but nobody ever thought that collations would not be an invasive
patch...
All expression nodes now have an explicit output-collation field, unless
they are known to only return a noncollatable data type (such as boolean
or record). Also, nodes that can invoke collation-aware functions store
a separate field that is the collation value to pass to the function.
This avoids confusion that arises when a function has collatable inputs
and noncollatable output type, or vice versa.
Also, replace the parser's on-the-fly collation assignment method with
a post-pass over the completed expression tree. This allows us to use
a more complex (and hopefully more nearly spec-compliant) assignment
rule without paying for it in extra storage in every expression node.
Fix assorted bugs in the planner's handling of collations by making
collation one of the defining properties of an EquivalenceClass and
by converting CollateExprs into discardable RelabelType nodes during
expression preprocessing.
This adds collation support for columns and domains, a COLLATE clause
to override it per expression, and B-tree index support.
Peter Eisentraut
reviewed by Pavel Stehule, Itagaki Takahiro, Robert Haas, Noah Misch
The original coding in tuplestore_trim() was only meant to work efficiently
in cases where each trim call deleted most of the tuples in the store.
Which, in fact, was the pattern of the original usage with a Material node
supporting mark/restore operations underneath a MergeJoin. However,
WindowAgg now uses tuplestores and it has considerably less friendly
trimming behavior. In particular it can attempt to trim one tuple at a
time off a large tuplestore. tuplestore_trim() had O(N^2) runtime in this
situation because of repeatedly shifting its tuple pointer array. Fix by
avoiding shifting the array until a reasonably large number of tuples have
been deleted. This can waste some pointer space, but we do still reclaim
the tuples themselves, so the percentage wastage should be pretty small.
Per Jie Li's report of slow percent_rank() evaluation. cume_dist() and
ntile() would certainly be affected as well, along with any other window
function that has a moving frame start and requires reading substantially
ahead of the current row.
Back-patch to 8.4, where window functions were introduced. There's no
need to tweak it before that.
Use a macro LogicalTapeReadExact() to encapsulate the error check when
we want to read an exact number of bytes from a "tape". Per a suggestion
of Takahiro Itagaki.
PL/pgSQL function within an exception handler. Make sure we use the right
resource owner when we create the tuplestore to hold returned tuples.
Simplify tuplestore API so that the caller doesn't need to be in the right
memory context when calling tuplestore_put* functions. tuplestore.c
automatically switches to the memory context used when the tuplestore was
created. Tuplesort was already modified like this earlier. This patch also
removes the now useless MemoryContextSwitch calls from callers.
Report by Aleksei on pgsql-bugs on Dec 22 2009. Backpatch to 8.1, like
the previous patch that broke this.
mode while callers hold pointers to in-memory tuples. I reported this for
the case of nodeWindowAgg's primary scan tuple, but inspection of the code
shows that all of the calls in nodeWindowAgg and nodeCtescan are at risk.
For the moment, fix it with a rather brute-force approach of copying
whenever one of the at-risk callers requests a tuple. Later we might
think of some sort of reference-count approach to reduce tuple copying.
some bufmgr probes, take out redundant and memory-leak-inducing path arguments
to smgr__md__read__done and smgr__md__write__done, fix bogus attempt to
recalculate space used in sort__done, clean up formatting in places where
I'm not sure pgindent will do a nice job by itself.
upcoming window-functions patch. First, tuplestore_trim is now an
exported function that must be explicitly invoked by callers at
appropriate times, rather than something that tuplestore tries to do
behind the scenes. Second, a read pointer that is marked as allowing
backward scan no longer prevents truncation. This means that a read pointer
marked as having BACKWARD but not REWIND capability can only safely read
backwards as far as the oldest other read pointer. (The expected use pattern
for this involves having another read pointer that serves as the truncation
fencepost.)
written to temp files by tuplesort.c and tuplestore.c. This saves 2 bytes per
row for 32-bit machines, and 6 bytes per row for 64-bit machines, which seems
worth the slight additional uglification of the tuple read/write routines.
There are some unimplemented aspects: recursive queries must use UNION ALL
(should allow UNION too), and we don't have SEARCH or CYCLE clauses.
These might or might not get done for 8.4, but even without them it's a
pretty useful feature.
There are also a couple of small loose ends and definitional quibbles,
which I'll send a memo about to pgsql-hackers shortly. But let's land
the patch now so we can get on with other development.
Yoshiyuki Asaba, with lots of help from Tatsuo Ishii and Tom Lane
This facility replaces the former mark/restore support but is otherwise
upward-compatible with previous uses. It's expected to be needed for
single evaluation of CTEs and also for window functions, so I'm committing
it separately instead of waiting for either one of those patches to be
finished. Per discussion with Greg Stark and Hitoshi Harada.
Note: I removed nodeFunctionscan's mark/restore support, instead of bothering
to update it for this change, because it was dead code anyway.
value. This means that hash index lookups are always lossy and have to be
rechecked when the heap is visited; however, the gain in index compactness
outweighs this when the indexed values are wide. Also, we only need to
perform datatype comparisons when the hash codes match exactly, rather than
for every entry in the hash bucket; so it could also win for datatypes that
have expensive comparison functions. A small additional win is gained by
keeping hash index pages sorted by hash code and using binary search to reduce
the number of index tuples we have to look at.
Xiao Meng
This commit also incorporates Zdenek Kotala's patch to isolate hash metapages
and hash bitmaps a bit better from the page header datastructures.
corresponding struct definitions. This allows other headers to avoid including
certain highly-loaded headers such as rel.h and relscan.h, instead using just
relcache.h, heapam.h or genam.h, which are more lightweight and thus cause less
unnecessary dependencies.
unnecessary #include lines in it. Also, move some tuple routine prototypes and
macros to htup.h, which allows removal of heapam.h inclusion from some .c
files.
For this to work, a new header file access/sysattr.h needed to be created,
initially containing attribute numbers of system columns, for pg_dump usage.
While at it, make contrib ltree, intarray and hstore header files more
consistent with our header style.
identical to tuplestore_puttuple(), except it operates on arrays of
Datums + nulls rather than a fully-formed HeapTuple. In several places
that use the tuplestore API, this means we can avoid creating a
HeapTuple altogether, saving a copy.
oprofile shows that a nontrivial amount of time is being spent in
repeated calls to index_getprocinfo, which really only needs to be
called once. So do that, and inline _hash_datum2hashkey to make it
work.
bucket number, so as to ensure locality of access to the index during the
insertion step. Without this, building an index significantly larger than
available RAM takes a very long time because of thrashing. On the other
hand, sorting is just useless overhead when the index does fit in RAM.
We choose to sort when the initial index size exceeds effective_cache_size.
This is a revised version of work by Tom Raney and Shreya Bhargava.
than dividing them into 1GB segments as has been our longtime practice. This
requires working support for large files in the operating system; at least for
the time being, it won't be the default.
Zdenek Kotala
regardless of the number of tuples involved, it's incorrect to skip it
when memtupcount = 1; the number of cycles saved is minuscule anyway.
An alternative solution would be to pull the state changes out to the
call site in tuplesort_performsort, but keeping them near the corresponding
changes in make_bounded_heap seems marginally cleaner. Noticed by
Greg Stark.
for each temp file, rather than once per sort or hashjoin; this allows
spreading the data of a large sort or join across multiple tablespaces.
(I remain dubious that this will make any difference in practice, but certain
people insisted.) Arrange to cache the results of parsing the GUC variable
instead of recomputing from scratch on every demand, and push usage of the
cache down to the bottommost fd.c level.
tablespace(s) in which to store temp tables and temporary files. This is a
list to allow spreading the load across multiple tablespaces (a random list
element is chosen each time a temp object is to be created). Temp files are
not stored in per-database pgsql_tmp/ directories anymore, but per-tablespace
directories.
Jaime Casanova and Albert Cervera, with review by Bernd Helmle and Tom Lane.
is using mark/restore but not rewind or backward-scan capability. Insert a
materialize plan node between a mergejoin and its inner child if the inner
child is a sort that is expected to spill to disk. The materialize shields
the sort from the need to do mark/restore and thereby allows it to perform
its final merge pass on-the-fly; while the materialize itself is normally
cheap since it won't spill to disk unless the number of tuples with equal
key values exceeds work_mem.
Greg Stark, with some kibitzing from Tom Lane.
need be returned. We keep a heap of the current best N tuples and sift-up
new tuples into it as we scan the input. For M input tuples this means
only about M*log(N) comparisons instead of M*log(M), not to mention a lot
less workspace when N is small --- avoiding spill-to-disk for large M
is actually the most attractive thing about it. Patch includes planner
and executor support for invoking this facility in ORDER BY ... LIMIT
queries. Greg Stark, with some editorialization by moi.
which comparison operators to use for plan nodes involving tuple comparison
(Agg, Group, Unique, SetOp). Formerly the executor looked up the default
equality operator for the datatype, which was really pretty shaky, since it's
possible that the data being fed to the node is sorted according to some
nondefault operator class that could have an incompatible idea of equality.
The planner knows what it has sorted by and therefore can provide the right
equality operator to use. Also, this change moves a couple of catalog lookups
out of the executor and into the planner, which should help startup time for
pre-planned queries by some small amount. Modify the planner to remove some
other cavalier assumptions about always being able to use the default
operators. Also add "nulls first/last" info to the Plan node for a mergejoin
--- neither the executor nor the planner can cope yet, but at least the API is
in place.
per-column options for btree indexes. The planner's support for this is still
pretty rudimentary; it does not yet know how to plan mergejoins with
nondefault ordering options. The documentation is pretty rudimentary, too.
I'll work on improving that stuff later.
Note incompatible change from prior behavior: ORDER BY ... USING will now be
rejected if the operator is not a less-than or greater-than member of some
btree opclass. This prevents less-than-sane behavior if an operator that
doesn't actually define a proper sort ordering is selected.
cases. Operator classes now exist within "operator families". While most
families are equivalent to a single class, related classes can be grouped
into one family to represent the fact that they are semantically compatible.
Cross-type operators are now naturally adjunct parts of a family, without
having to wedge them into a particular opclass as we had done originally.
This commit restructures the catalogs and cleans up enough of the fallout so
that everything still works at least as well as before, but most of the work
needed to actually improve the planner's behavior will come later. Also,
there are not yet CREATE/DROP/ALTER OPERATOR FAMILY commands; the only way
to create a new family right now is to allow CREATE OPERATOR CLASS to make
one by default. I owe some more documentation work, too. But that can all
be done in smaller pieces once this infrastructure is in place.
repeatedly. Now that we don't have to worry about memory leaks from
glibc's qsort, we can safely put CHECK_FOR_INTERRUPTS into the tuplesort
comparators, as was requested a couple months ago. Also, get rid of
non-reentrancy and an extra level of function call in tuplesort.c by
providing a variant qsort_arg() API that passes an extra void * argument
through to the comparison routine. (We might want to use that in other
places too, I didn't look yet.)
per-tuple space overhead for sorts in memory. I chose to replace the
previous patch that tried to write out the bare minimum amount of data
when sorting on disk; instead, just dump the MinimalTuples as-is. This
wastes 3 to 10 bytes per tuple depending on architecture and null-bitmap
length, but the simplification in the writetup/readtup routines seems
worth it.
tuples with less header overhead than a regular HeapTuple, per my
recent proposal. Teach TupleTableSlot code how to deal with these.
As proof of concept, change tuplestore.c to store MinimalTuples instead
of HeapTuples. Future patches will expand the concept to other places
where it is useful.
and transaction visibility fields of tuples being sorted. These are
always uninteresting in a tuple being sorted (if the fields were actually
selected, they'd have been pulled out into user columns beforehand).
This saves about 24 bytes per row being sorted, which is a useful savings
for any but the widest of sort rows. Per recent discussion.
case where we run low on array slots before we run low on memory is much
more probable than I had thought, and so it's important to treat each
tape fairly in that case. To fix this, track per-tape slot allocations
just like we track per-tape space allocation. Also, in the FINALMERGE
code path avoid scanning all the input tapes when we really only need to
read from one. This should fix poor behavior with very large work_mem
as exhibited by Stefan Kaltenbrunner.
I didn't do anything about putting an upper bound on the number of tapes,
but maybe we should still consider that.
performance issue during regular merge passes not only the 'final merge'
case. The original design contemplated that there would never be more
than about one free block per 'tape', hence no need for an efficient
method of keeping the free blocks sorted. But given the later addition
of merge preread behavior in tuplesort.c, there is likely to be about
work_mem worth of free blocks, which is not so small ... and for that
matter the number of tapes isn't necessarily small anymore either. So
we'd better get rid of the assumption entirely. Instead, I'm assuming
that the usage pattern will involve alternation between merge preread
and writing of a new run. This makes it reasonable to just add blocks
to the list without sorting during successive ltsReleaseBlock calls,
and then do a qsort() when we start getting ltsGetFreeBlock() calls.
Experimentation seems to confirm that there aren't many qsort calls
relative to the number of ltsReleaseBlock/ltsGetFreeBlock calls.
we are doing the final merge pass on-the-fly, and not writing the data
back onto a 'tape', the number of free blocks in the tape set will become
large, leading to a lot of time wasted in ltsReleaseBlock(). There is
really no need to track the free blocks anymore in this state, so add a
simple shutoff switch. Per report from Stefan Kaltenbrunner.
In particular, ensure that enlargement of the memtuples[] array doesn't
fall foul of MaxAllocSize when work_mem is very large, and don't bother
enlarging it if that would force an immediate switch into 'tape' mode anyway.
we'll go over to disk-based sort if we reach that limit.
This fixes Stefan Kaltenbrunner's observation that sorting can suffer an
'invalid memory alloc request size' failure when sort_mem is set large
enough. It's unfortunately not so easy to fix in 8.1 ...
each tuple, as per my proposal of several days ago. Also, clean up
sort memory management by keeping all working data in a separate memory
context, and refine the handling of low-memory conditions.
allocates the control data. The per-tape buffers are allocated only
on first use. This saves memory in situations where tuplesort.c
overestimates the number of tapes needed (ie, there are fewer runs
than tapes). Also, this makes legitimate the coding in inittapes()
that includes tape buffer space in the maximum-memory calculation:
when inittapes runs, we've already expended the whole allowed memory
on tuple storage, and so we'd better not allocate all the tape buffers
until we've flushed some tuples out of memory.
with fixed merge order (fixed number of "tapes") was based on obsolete
assumptions, namely that tape drives are expensive. Since our "tapes"
are really just a couple of buffers, we can have a lot of them given
adequate workspace. This allows reduction of the number of merge passes
with consequent savings of I/O during large sorts.
Simon Riggs with some rework by Tom Lane
comment line where output as too long, and update typedefs for /lib
directory. Also fix case where identifiers were used as variable names
in the backend, but as typedefs in ecpg (favor the backend for
indenting).
Backpatch to 8.1.X.
which is neither needed by nor related to that header. Remove the bogus
inclusion and instead include the header in those C files that actually
need it. Also fix unnecessary inclusions and bad inclusion order in
tsearch2 files.
Also performed an initial run through of upgrading our Copyright date to
extend to 2005 ... first run here was very simple ... change everything
where: grep 1996-2004 && the word 'Copyright' ... scanned through the
generated list with 'less' first, and after, to make sure that I only
picked up the right entries ...
TID (heap position). This doesn't do anything to the validity of the
finished index, but by pretending to qsort() that there are no really
equal keys in the sort, we can avoid performance problems with qsort
implementations that have trouble with large numbers of equal keys.
Patch from Manfred Koizar.
Make btree index creation and initial validation of foreign-key constraints
use maintenance_work_mem rather than work_mem as their memory limit.
Add some code to guc.c to allow these variables to be referenced by their
old names in SHOW and SET commands, for backwards compatibility.
pghackers proposal of 8-Nov. All the existing cross-type comparison
operators (int2/int4/int8 and float4/float8) have appropriate support.
The original proposal of storing the right-hand-side datatype as part of
the primary key for pg_amop and pg_amproc got modified a bit in the event;
it is easier to store zero as the 'default' case and only store a nonzero
when the operator is actually cross-type. Along the way, remove the
long-since-defunct bigbox_ops operator class.
Remove the 'strategy map' code, which was a large amount of mechanism
that no longer had any use except reverse-mapping from procedure OID to
strategy number. Passing the strategy number to the index AM in the
first place is simpler and faster.
This is a preliminary step in planned support for cross-datatype index
operations. I'm committing it now since the ScanKeyEntryInitialize()
API change touches quite a lot of files, and I want to commit those
changes before the tree drifts under me.
datatype by array_eq and array_cmp; use this to solve problems with memory
leaks in array indexing support. The parser's equality_oper and ordering_oper
routines also use the cache. Change the operator search algorithms to look
for appropriate btree or hash index opclasses, instead of assuming operators
named '<' or '=' have the right semantics. (ORDER BY ASC/DESC now also look
at opclasses, instead of assuming '<' and '>' are the right things.) Add
several more index opclasses so that there is no regression in functionality
for base datatypes. initdb forced due to catalog additions.
This makes no difference for existing uses, but allows SelectSortFunction()
and pred_test_simple_clause() to use indexscans instead of seqscans to
locate entries for a particular operator in pg_amop. Better yet, they can
use the SearchSysCacheList() API to cache the search results.
(materialization into a tuple store) discussed on pgsql-hackers earlier.
I've updated the documentation and the regression tests.
Notes on the implementation:
- I needed to change the tuple store API slightly -- it assumes that it
won't be used to hold data across transaction boundaries, so the temp
files that it uses for on-disk storage are automatically reclaimed at
end-of-transaction. I added a flag to tuplestore_begin_heap() to control
this behavior. Is changing the tuple store API in this fashion OK?
- in order to store executor results in a tuple store, I added a new
CommandDest. This works well for the most part, with one exception: the
current DestFunction API doesn't provide enough information to allow the
Executor to store results into an arbitrary tuple store (where the
particular tuple store to use is chosen by the call site of
ExecutorRun). To workaround this, I've temporarily hacked up a solution
that works, but is not ideal: since the receiveTuple DestFunction is
passed the portal name, we can use that to lookup the Portal data
structure for the cursor and then use that to get at the tuple store the
Portal is using. This unnecessarily ties the Portal code with the
tupleReceiver code, but it works...
The proper fix for this is probably to change the DestFunction API --
Tom suggested passing the full QueryDesc to the receiveTuple function.
In that case, callers of ExecutorRun could "subclass" QueryDesc to add
any additional fields that their particular CommandDest needed to get
access to. This approach would work, but I'd like to think about it for
a little bit longer before deciding which route to go. In the mean time,
the code works fine, so I don't think a fix is urgent.
- (semi-related) I added a NO SCROLL keyword to DECLARE CURSOR, and
adjusted the behavior of SCROLL in accordance with the discussion on
-hackers.
- (unrelated) Cleaned up some SGML markup in sql.sgml, copy.sgml
Neil Conway
entire contents of the subplan into the tuplestore before we can return
any tuples. Instead, the tuplestore holds what we've already read, and
we fetch additional rows from the subplan as needed. Random access to
the previously-read rows works with the tuplestore, and doesn't affect
the state of the partially-read subplan. This is a step towards fixing
the problems with cursors over complex queries --- we don't want to
stick in Materialize nodes if they'll prevent quick startup for a cursor.
just the significant fields of FunctionCallInfoData, rather than MemSet'ing
the whole struct to zero. Unused positions in the arg[] array will
thereby contain garbage rather than zeroes. This buys back some of the
performance hit from increasing FUNC_MAX_ARGS. Also tweak tuplesort.c
code for more speed by marking some routines 'inline'. All together
these changes speed up simple sorts, like count(distinct int4column),
by about 25% on a P4 running RH Linux 7.2.
value '-2' is used to indicate a variable-width type whose width is
computed as strlen(datum)+1. Everything that looks at typlen is updated
except for array support, which Joe Conway is working on; at the moment
it wouldn't work to try to create an array of cstring.
to make a reasonable attempt at accounting for palloc overhead, not just
the requested size of each memory chunk. Since in many scenarios this
will make for a significant reduction in the amount of space acquired,
partially compensate by doubling the default value of SORT_MEM to 1Mb.
Per discussion in pgsql-general around 9-Jun-2002..
yesterday's proposal to pghackers. Also remove unnecessary parameters
to heap_beginscan, heap_rescan. I modified pg_proc.h to reflect the
new numbers of parameters for the AM interface routines, but did not
force an initdb because nothing actually looks at those fields.
pgsql-hackers. pg_opclass now has a row for each opclass supported by each
index AM, not a row for each opclass name. This allows pg_opclass to show
directly whether an AM supports an opclass, and furthermore makes it possible
to store additional information about an opclass that might be AM-dependent.
pg_opclass and pg_amop now store "lossy" and "haskeytype" information that we
previously expected the user to remember to provide in CREATE INDEX commands.
Lossiness is no longer an index-level property, but is associated with the
use of a particular operator in a particular index opclass.
Along the way, IndexSupportInitialize now uses the syscaches to retrieve
pg_amop and pg_amproc entries. I find this reduces backend launch time by
about ten percent, at the cost of a couple more special cases in catcache.c's
IndexScanOK.
Initial work by Oleg Bartunov and Teodor Sigaev, further hacking by Tom Lane.
initdb forced.
(vs. at the end of a normal sort). This ensures that explicit sorts
yield the same ordering as a btree index scan. To be really sure that
that equivalence holds, we use the btree entries in pg_amop to decide
whether we are looking at a '<' or '>' operator. For a sort operator
that has no btree association, we put the nulls at the front if the
operator is named '>' ... pretty grotty, but it does the right thing in
simple ASC and DESC cases, and at least there's no possibility of getting
a different answer depending on the plan type chosen.
a separate statement (though it can still be invoked as part of VACUUM, too).
pg_statistic redesigned to be more flexible about what statistics are
stored. ANALYZE now collects a list of several of the most common values,
not just one, plus a histogram (not just the min and max values). Random
sampling is used to make the process reasonably fast even on very large
tables. The number of values and histogram bins collected is now
user-settable via an ALTER TABLE command.
There is more still to do; the new stats are not being used everywhere
they could be in the planner. But the remaining changes for this project
should be localized, and the behavior is already better than before.
A not-very-related change is that sorting now makes use of btree comparison
routines if it can find one, rather than invoking '<' twice.
allocated by plan nodes are not leaked at end of query. This doesn't
really matter for normal queries, but it sure does for queries invoked
repetitively inside SQL functions. Clean up some other grotty code
associated with tupdescs, and fix a few other memory leaks exposed by
tests with simple SQL functions.
maintained for each cache entry. A cache entry will not be freed until
the matching ReleaseSysCache call has been executed. This eliminates
worries about cache entries getting dropped while still in use. See
my posting to pg-hackers of even date for more info.
materialized tupleset is small enough) instead of a temporary relation.
This was something I was thinking of doing anyway for performance, and Jan
says he needs it for TOAST because he doesn't want to cope with toasting
noname relations. With this change, the 'noname table' support in heap.c
is dead code, and I have accordingly removed it. Also clean up 'noname'
plan handling in planner --- nonames are either sort or materialize plans,
and it seems less confusing to handle them separately under those names.
Robert Haas
Heikki Linnakangas
Tom Lane
Bruce Momjian
Teodor Sigaev
Andres Freund
Stephen Frost
Peter Eisentraut
Noah Misch
Alvaro Herrera
Magnus Hagander
Neil Conway
PostgreSQL Daemon