to 'Size' (that is, size_t), and install overflow detection checks in it.
This allows us to remove the former arbitrary restrictions on NBuffers
etc. It won't make any difference in a 32-bit machine, but in a 64-bit
machine you could theoretically have terabytes of shared buffers.
(How efficiently we could manage 'em remains to be seen.) Similarly,
num_temp_buffers, work_mem, and maintenance_work_mem can be set above
2Gb on a 64-bit machine. Original patch from Koichi Suzuki, additional
work by moi.
insufficient paranoia in code that follows t_ctid links. (We must do both
because even with VACUUM doing it properly, the intermediate state with
a dangling t_ctid link is visible concurrently during lazy VACUUM, and
could be seen afterwards if either type of VACUUM crashes partway through.)
Also try to improve documentation about what's going on. Patch is a bit
bulky because passing the XMAX information around required changing the
APIs of some low-level heapam.c routines, but it's not conceptually very
complicated. Per trouble report from Teodor and subsequent analysis.
This needs to be back-patched, but I'll do that after 8.1 beta is out.
whenever we generate a new OID. This prevents occasional duplicate-OID
errors that can otherwise occur once the OID counter has wrapped around.
Duplicate relfilenode values are also checked for when creating new
physical files. Per my recent proposal.
delay and limit, both as global GUCs and as table-specific entries in
pg_autovacuum. stats_reset_on_server_start is now OFF by default,
but a reset is forced if we did WAL replay. XID-wrap vacuums do not
ANALYZE, but do FREEZE if it's a template database. Alvaro Herrera
ResourceOwner mechanism already released all reference counts for the
cache entries; therefore, we do not need to scan the catcache or relcache
at transaction end, unless we want to do it as a debugging crosscheck.
Do the crosscheck only in Assert mode. This is the same logic we had
previously installed in AtEOXact_Buffers to avoid overhead with large
numbers of shared buffers. I thought it'd be a good idea to do it here
too, in view of Kari Lavikka's recent report showing a real-world case
where AtEOXact_CatCache is taking a significant fraction of runtime.
track shared relations in a separate hashtable, so that operations done
from different databases are counted correctly. Add proper support for
anti-XID-wraparound vacuuming, even in databases that are never connected
to and so have no stats entries. Miscellaneous other bug fixes.
Alvaro Herrera, some additional fixes by Tom Lane.
Also, write multiple WAL buffers out in one write() operation.
ITAGAKI Takahiro
---------------------------------------------------------------------------
> If we disable writeback-cache and use open_sync, the per-page writing
> behavior in WAL module will show up as bad result. O_DIRECT is similar
> to O_DSYNC (at least on linux), so that the benefit of it will disappear
> behind the slow disk revolution.
>
> In the current source, WAL is written as:
> for (i = 0; i < N; i++) { write(&buffers[i], BLCKSZ); }
> Is this intentional? Can we rewrite it as follows?
> write(&buffers[0], N * BLCKSZ);
>
> In order to achieve it, I wrote a 'gather-write' patch (xlog.gw.diff).
> Aside from this, I'll also send the fixed direct io patch (xlog.dio.diff).
> These two patches are independent, so they can be applied either or both.
>
>
> I tested them on my machine and the results as follows. It shows that
> direct-io and gather-write is the best choice when writeback-cache is off.
> Are these two patches worth trying if they are used together?
>
>
> | writeback | fsync= | fdata | open_ | fsync_ | open_
> patch | cache | false | sync | sync | direct | direct
> ------------+-----------+--------+-------+-------+--------+---------
> direct io | off | 124.2 | 105.7 | 48.3 | 48.3 | 48.2
> direct io | on | 129.1 | 112.3 | 114.1 | 142.9 | 144.5
> gather-write| off | 124.3 | 108.7 | 105.4 | (N/A) | (N/A)
> both | off | 131.5 | 115.5 | 114.4 | 145.4 | 145.2
>
> - 20runs * pgbench -s 100 -c 50 -t 200
> - with tuning (wal_buffers=64, commit_delay=500, checkpoint_segments=8)
> - using 2 ATA disks:
> - hda(reiserfs) includes system and wal.
> - hdc(jfs) includes database files. writeback-cache is always on.
>
> ---
> ITAGAKI Takahiro
chdir into PGDATA and subsequently use relative paths instead of absolute
paths to access all files under PGDATA. This seems to give a small
performance improvement, and it should make the system more robust
against naive DBAs doing things like moving a database directory that
has a live postmaster in it. Per recent discussion.
basic regression tests for GiST to the standard regression tests.
I took the opportunity to add an rtree-equivalent gist opclass for
circles; the contrib version only covered boxes and polygons, but
indexing circles is very handy for distance searches.
the difference between checkpoints forced due to WAL segment consumption
and checkpoints forced for other reasons (such as CREATE DATABASE). Avoid
generating 'checkpoints are occurring too frequently' messages when the
checkpoint wasn't caused by WAL segment consumption. Per gripe from
Chris K-L.
current time: provide a GetCurrentTimestamp() function that returns
current time in the form of a TimestampTz, instead of separate time_t
and microseconds fields. This is what all the callers really want
anyway, and it eliminates low-level dependencies on AbsoluteTime,
which is a deprecated datatype that will have to disappear eventually.
and pg_auth_members. There are still many loose ends to finish in this
patch (no documentation, no regression tests, no pg_dump support for
instance). But I'm going to commit it now anyway so that Alvaro can
make some progress on shared dependencies. The catalog changes should
be pretty much done.
- full concurrency for insert/update/select/vacuum:
- select and vacuum never locks more than one page simultaneously
- select (gettuple) hasn't any lock across it's calls
- insert never locks more than two page simultaneously:
- during search of leaf to insert it locks only one page
simultaneously
- while walk upward to the root it locked only parent (may be
non-direct parent) and child. One of them X-lock, another may
be S- or X-lock
- 'vacuum full' locks index
- improve gistgetmulti
- simplify XLOG records
Fix bug in index_beginscan_internal: LockRelation may clean
rd_aminfo structure, so move GET_REL_PROCEDURE after LockRelation
to the existing X-direction tests. An rtree class now includes 4 actual
2-D tests, 4 1-D X-direction tests, and 4 1-D Y-direction tests.
This involved adding four new Y-direction test operators for each of
box and polygon; I followed the PostGIS project's lead as to the names
of these operators.
NON BACKWARDS COMPATIBLE CHANGE: the poly_overleft (&<) and poly_overright
(&>) operators now have semantics comparable to box_overleft and box_overright.
This is necessary to make r-tree indexes work correctly on polygons.
Also, I changed circle_left and circle_right to agree with box_left and
box_right --- formerly they allowed the boundaries to touch. This isn't
actually essential given the lack of any r-tree opclass for circles, but
it seems best to sync all the definitions while we are at it.
polygon operators (<<, &<, >>, &>). Per ideas originally put forward
by andrew@supernews and later rediscovered by moi. This patch just
fixes the existing opclasses, and does not add any new behavior as I
proposed earlier; that can be sorted out later. In principle this
could be back-patched, since it changes only search behavior and not
system catalog entries nor rtree index contents. I'm not currently
planning to do that, though, since I think it could use more testing.
(a/k/a SELECT INTO). Instead, flush and fsync the whole relation before
committing. We do still need the WAL log when PITR is active, however.
Simon Riggs and Tom Lane.
2. improve vacuum for gist
- use FSM
- full vacuum:
- reforms parent tuple if it's needed
( tuples was deleted on child page or parent tuple remains invalid
after crash recovery )
- truncate index file if possible
3. fixes bugs and mistakes
scankeys arrays that it needs can never have more than INDEX_MAX_KEYS
entries, so it's reasonable to just allocate them as fixed-size local
arrays, and save the cost of palloc/pfree. Not a huge savings, but
a cycle saved is a cycle earned ...
includes error checking and an appropriate ereport(ERROR) message.
This gets rid of rather tedious and error-prone manipulation of errno,
as well as a Windows-specific bug workaround, at more than a dozen
call sites. After an idea in a recent patch by Heikki Linnakangas.
given reasonably short lifespans for prepared transactions, this should
mean that only a small minority of state files ever need to be fsynced
at all. Per discussion with Heikki Linnakangas.
old suggestion by Oliver Jowett. Also, add a transaction column to the
pg_locks view to show the xid of each transaction holding or awaiting
locks; this allows prepared transactions to be properly associated with
the locks they own. There was already a column named 'transaction',
and I chose to rename it to 'transactionid' --- since this column is
new in the current devel cycle there should be no backwards compatibility
issue to worry about.
nonconsecutive columns of a multicolumn index, as per discussion around
mid-May (pghackers thread "Best way to scan on-disk bitmaps"). This
turns out to require only minimal changes in btree, and so far as I can
see none at all in GiST. btcostestimate did need some work, but its
original assumption that index selectivity == heap selectivity was
quite bogus even before this.
transaction IDs, rather than like subtrans; in particular, the information
now survives a database restart. Per previous discussion, this is
essential for PITR log shipping and for 2PC.
up have the standard layout with unused space between pd_lower and pd_upper.
When this is set, XLogInsert will omit the unused space without bothering
to scan it to see if it's zero. That saves time in XLogInsert, and also
allows reversion of my earlier patch to make PageRepairFragmentation et al
explicitly re-zero freed space. Per suggestion by Heikki Linnakangas.
Instead of a separate CRC on each backup block, include backup blocks
in their parent WAL record's CRC; this is important to ensure that the
backup block really goes with the WAL record, ie there was not a page
tear right at the start of the backup block. Implement a simple form
of compression of backup blocks: drop any run of zeroes starting at
pd_lower, so as not to store the unused 'hole' that commonly exists in
PG heap and index pages. Tweak PageRepairFragmentation and related
routines to ensure they keep the unused space zeroed, so that the above
compression method remains effective. All per recent discussions.
spotted by Qingqing Zhou. The HASH_ENTER action now automatically
fails with elog(ERROR) on out-of-memory --- which incidentally lets
us eliminate duplicate error checks in quite a bunch of places. If
you really need the old return-NULL-on-out-of-memory behavior, you
can ask for HASH_ENTER_NULL. But there is now an Assert in that path
checking that you aren't hoping to get that behavior in a palloc-based
hash table.
Along the way, remove the old HASH_FIND_SAVE/HASH_REMOVE_SAVED actions,
which were not being used anywhere anymore, and were surely too ugly
and unsafe to want to see revived again.
routines in the index's relcache entry, instead of doing a fresh fmgr_info
on every index access. We were already doing this for the index's opclass
support functions; not sure why we didn't think to do it for the AM
functions too. This supersedes the former method of caching (only)
amgettuple in indexscan scan descriptors; it's an improvement because the
function lookup can be amortized across multiple statements instead of
being repeated for each statement. Even though lookup for builtin
functions is pretty cheap, this seems to drop a percent or two off some
simple benchmarks.
communication structure, and make it its own module with its own lock.
This should reduce contention at least a little, and it definitely makes
the code seem cleaner. Per my recent proposal.
external projects, we should be careful about what parts of the GiST
API are considered implementation details, and which are part of the
public API. Therefore, I've moved internal-only declarations into
gist_private.h -- future backward-incompatible changes to gist.h should
be made with care, to avoid needlessly breaking external GiST extensions.
Also did some related header cleanup: remove some unnecessary #includes
from gist.h, and remove some unused definitions: isAttByVal(), _gistdump(),
and GISTNStrategies.
- make sure we always invoke user-supplied GiST methods in a short-lived
memory context. This means the backend isn't exposed to any memory leaks
that be in those methods (in fact, it is probably a net loss for most
GiST methods to bother manually freeing memory now). This also means
we can do away with a lot of ugly manual memory management in the
GiST code itself.
- keep the current page of a GiST index scan pinned, rather than doing a
ReadBuffer() for each tuple produced by the scan. Since ReadBuffer() is
expensive, this is a perf. win
- implement dead tuple killing for GiST indexes (which is easy to do, now
that we keep a pin on the current scan page). Now all the builtin indexes
implement dead tuple killing.
- cleanup a lot of ugly code in GiST
in an inconsistent state. (This is only latent because in reality
ExecSeqRestrPos is dead code at the moment ... but someday maybe it won't
be.) Add some comments about what the API for plan node mark/restore
actually is, because it's not immediately obvious.
methods: they all invoke UpdateStats() since they have computed the
number of heap tuples, so I created a function in catalog/index.c that
each AM now calls.
memset() or MemSet() to a char *. For one, memset()'s first argument is
a void *, and further void * can be implicitly coerced to/from any other
pointer type.
and VACUUM: in the interval between adding a new page to the relation
and formatting it, it was possible for VACUUM to come along and decide
it should format the page too. Though not harmful in itself, this would
cause data loss if a third transaction were able to insert tuples into
the vacuumed page before the original extender got control back.
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.
only one argument. (Per recent discussion, the option to accept multiple
arguments is pretty useless for user-defined types, and would be a likely
source of security holes if it was used.) Simplify call sites of
output/send functions to not bother passing more than one argument.
is contention for a tuple-level lock. This solves the problem of a
would-be exclusive locker being starved out by an indefinite succession
of share-lockers. Per recent discussion with Alvaro.
Essentially, we shoehorn in a lockable-object-type field by taking
a byte away from the lockmethodid, which can surely fit in one byte
instead of two. This allows less artificial definitions of all the
other fields of LOCKTAG; we can get rid of the special pg_xactlock
pseudo-relation, and also support locks on individual tuples and
general database objects (including shared objects). None of those
possibilities are actually exploited just yet, however.
I removed pg_xactlock from pg_class, but did not force initdb for
that change. At this point, relkind 's' (SPECIAL) is unused and
could be removed entirely.
to eliminate unnecessary deadlocks. This commit adds SELECT ... FOR SHARE
paralleling SELECT ... FOR UPDATE. The implementation uses a new SLRU
data structure (managed much like pg_subtrans) to represent multiple-
transaction-ID sets. When more than one transaction is holding a shared
lock on a particular row, we create a MultiXactId representing that set
of transactions and store its ID in the row's XMAX. This scheme allows
an effectively unlimited number of row locks, just as we did before,
while not costing any extra overhead except when a shared lock actually
has to be shared. Still TODO: use the regular lock manager to control
the grant order when multiple backends are waiting for a row lock.
Alvaro Herrera and Tom Lane.
critical and noncritical contexts (an example of noncritical being
post-checkpoint removal of dead xlog segments). In the critical cases
the CRIT_SECTION mechanism will cause ERROR to be promoted to PANIC
anyway, and in the noncritical cases we shouldn't let an error take
down the entire database. Arguably there should be *no* explicit
PANIC errors in this module, only more START/END_CRIT_SECTION calls,
but I didn't go that far. (Yet.)
when recycling a large number of xlog segments during checkpoint.
The former behavior searched from the same start point each time,
requiring O(checkpoint_segments^2) stat() calls to relocate all the
segments. Instead keep track of where we stopped last time through.
assuming comparison of atttypid is sufficient. In a dropped column
atttypid will be 0, and we'd better check the physical-storage data
to make sure the tupdescs are physically compatible.
I do not believe there is a real risk before 8.0, since before that
we only used this routine to compare successive states of the tupdesc
for a particular relation. But 8.0's typcache.c might be comparing
arbitrary tupdescs so we'd better play it safer.
indexes. Replace all heap_openr and index_openr calls by heap_open
and index_open. Remove runtime lookups of catalog OID numbers in
various places. Remove relcache's support for looking up system
catalogs by name. Bulky but mostly very boring patch ...
avoid encroaching on the 'user' range of OIDs by allowing automatic
OID assignment to use values below 16k until we reach normal operation.
initdb not forced since this doesn't make any incompatible change;
however a lot of stuff will have different OIDs after your next initdb.
deferred triggers: either one can create more work for the other,
so we have to loop till it's all gone. Per example from andrew@supernews.
Add a regression test to help spot trouble in this area in future.
few palloc's. I also chose to eliminate the restype and restypmod fields
entirely, since they are redundant with information stored in the node's
contained expression; re-examining the expression at need seems simpler
and more reliable than trying to keep restype/restypmod up to date.
initdb forced due to change in contents of stored rules.
change saves a great deal of space in pg_proc and its primary index,
and it eliminates the former requirement that INDEX_MAX_KEYS and
FUNC_MAX_ARGS have the same value. INDEX_MAX_KEYS is still embedded
in the on-disk representation (because it affects index tuple header
size), but FUNC_MAX_ARGS is not. I believe it would now be possible
to increase FUNC_MAX_ARGS at little cost, but haven't experimented yet.
There are still a lot of vestigial references to FUNC_MAX_ARGS, which
I will clean up in a separate pass. However, getting rid of it
altogether would require changing the FunctionCallInfoData struct,
and I'm not sure I want to buy into that.
transaction rollback via UNDO but I think that's highly unlikely to
happen, so we may as well remove the stubs. (Someday we ought to
rip out the stub xxx_undo routines, too.) Per Alvaro.
access: define new index access method functions 'amgetmulti' that can
fetch multiple TIDs per call. (The functions exist but are totally
untested as yet.) Since I was modifying pg_am anyway, remove the
no-longer-needed 'rel' parameter from amcostestimate functions, and
also remove the vestigial amowner column that was creating useless
work for Alvaro's shared-object-dependencies project.
Initdb forced due to changes in pg_am.
currently does. This is now the default Win32 wal sync method because
we perfer o_datasync to fsync.
Also, change Win32 fsync to a new wal sync method called
fsync_writethrough because that is the behavior of _commit, which is
what is used for fsync on Win32.
Backpatch to 8.0.X.
PageIndexTupleDelete() with a single pass of compactification ---
logic mostly lifted from PageRepairFragmentation. I noticed while
profiling that a VACUUM that's cleaning up a whole lot of deleted
tuples would spend as much as a third of its CPU time in
PageIndexTupleDelete; not too surprising considering the loop method
was roughly O(N^2) in the number of tuples involved.