messages to the stats collector. This avoids the problem that enabling
stats_row_level for autovacuum has a significant overhead for short
read-only transactions, as noted by Arjen van der Meijden. We can avoid
an extra gettimeofday call by piggybacking on the one done for WAL-logging
xact commit or abort (although that doesn't help read-only transactions,
since they don't WAL-log anything).
In my proposal for this, I noted that we could change the WAL log entries
for commit/abort to record full TimestampTz precision, instead of only
time_t as at present. That's not done in this patch, but will be committed
separately.
when a relation is opened multiple times in the same transaction. This is
particularly useful for system catalogs, which we may heap_open or index_open
many times in a transaction, and it doesn't really cost anything extra even
if the rel is touched but once. Motivated by study of an example from Greg
Stark, in which pgstat_initstats() accounted for an unreasonably large
fraction of the runtime.
its table list and then rechecks pgstat before vacuuming each table to
verify that no one has vacuumed the table in the meantime.
In the current autovacuum world this only means that a worker will not
vacuum a table that a user has vacuumed manually after the worker started.
When support for multiple autovacuum workers is introduced, this will reduce
the probability of simultaneous workers on the same database doing redundant
work.
continuously, and requests vacuum runs of "autovacuum workers" to postmaster.
The workers do the actual vacuum work. This allows for future improvements,
like allowing multiple autovacuum jobs running in parallel.
For now, the code keeps the original behavior of having a single autovac
process at any time by sleeping until the previous worker has finished.
entries for the victim database go away sooner rather than later. We already
did the equivalent thing at the per-relation level, not sure why it's not
been done for whole databases. With this change, pgstat_vacuum_tabstat
should usually not find anything to do; though we still need it as a backstop
in case DROPDB or TABPURGE messages get lost under load.
already collected in the current transaction; this allows plpgsql functions to
watch for stats updates even though they are confined to a single transaction.
Use this instead of the previous kluge involving pg_stat_file() to wait for
the stats collector to update in the stats regression test. Internally,
decouple storage of stats snapshots from transaction boundaries; they'll
now stick around until someone calls pgstat_clear_snapshot --- which xact.c
still does at transaction end, to maintain the previous behavior. This makes
the logic a lot cleaner, at the price of a couple dozen cycles per transaction
exit.
input in the stats collector. Our select() emulation is apparently buggy
for UDP sockets :-(. This should resolve problems with stats collection
(and hence autovacuum) failing under more than minimal load. Diagnosis
and patch by Magnus Hagander.
Patch probably needs to be back-ported to 8.1 and 8.0, but first let's
see if it makes the buildfarm happy...
(or other types of pg_class entry): the function pgstat_vacuum_tabstat,
invoked during VACUUM startup, had runtime proportional to the number of
stats table entries times the number of pg_class rows; in other words
O(N^2) if the stats collector's information is reasonably complete.
Replace list searching with a hash table to bring it back to O(N)
behavior. Per report from kim at myemma.com.
Back-patch as far as 8.1; 8.0 and before use different coding here.
identify long-running transactions. Since we already need to record
the transaction-start time (e.g. for now()), we don't need any
additional system calls to report this information.
Catversion bumped, initdb required.
Windows), arrange for each postmaster child process to be its own process
group leader, and deliver signals SIGINT, SIGTERM, SIGQUIT to the whole
process group not only the direct child process. This provides saner behavior
for archive and recovery scripts; in particular, it's possible to shut down a
warm-standby recovery server using "pg_ctl stop -m immediate", since delivery
of SIGQUIT to the startup subprocess will result in killing the waiting
recovery_command. Also, this makes Query Cancel and statement_timeout apply
to scripts being run from backends via system(). (There is no support in the
core backend for that, but it's widely done using untrusted PLs.) Per gripe
from Stephen Harris and subsequent discussion.
pgstat_bestart() has been called; else any lock-block occurring
during InitPostgres() is disastrous. I believe this explains
recent wasp regression failure; at least it explains the crash I
got while trying to duplicate the problem. I also made
pgstat_report_activity() safe against the same scenario, just
in case. The report_waiting hazard was created by my patch of
19-Aug to include waiting status in pg_stat_activity.
that ps_status provides by appending 'waiting' to the PS display. This
completes the project of making it feasible to turn off process title
updates and instead rely on pg_stat_activity. Per my suggestion a few
weeks ago.
EINTR; the stats code was failing to do this and so were a couple of places
in the postmaster. The stats code assumed that recv() could not return EINTR
if a preceding select() showed the socket to be read-ready, but this is
demonstrably false with our Windows implementation of recv(), and it may
not be the case on all Unix variants either. I think this explains the
intermittent stats regression test failures we've been seeing, as well
as reports of stats collector instability under high load on Windows.
Backpatch as far as 8.0.
be delivered directly to the collector process. The extra process context
swaps required to transfer data through the buffer process seem to outweigh
any value the buffering might have. Per recent discussion and tests.
I modified Bruce's draft patch to use poll() rather than select() where
available (this makes a noticeable difference on my system), and fixed
up the EXEC_BACKEND case.
analyzing, so that future analyze threshold calculations don't get confused.
Also, make sure we correctly track the decrease of live tuples cause by
deletes.
Per report from Dylan Hansen, patches by Tom Lane and me.
changing semantics too much. statement_timestamp is now set immediately
upon receipt of a client command message, and the various places that used
to do their own gettimeofday() calls to mark command startup are referenced
to that instead. I have also made stats_command_string use that same
value for pg_stat_activity.query_start for both the command itself and
its eventual replacement by <IDLE> or <idle in transaction>. There was
some debate about that, but no argument that seemed convincing enough to
justify an extra gettimeofday() call.
current commands; instead, store current-status information in shared
memory. This substantially reduces the overhead of stats_command_string
and also ensures that pg_stat_activity is fully up to date at all times.
Per my recent proposal.
---------------------------------------------------------------------------
Delay write of pg_stats file to once every five minutes, during
shutdown, or when requested by a backend:
It changes so the file is only written once every 5 minutes (changeable
of course, I just picked something) instead of once every half second.
It's still written when the stats collector shuts down, just as before.
And it is now also written on backend request. A backend requests a
rewrite by simply sending a special stats message. It operates on the
assumption that the backends aren't actually going to read the
statistics file very often, compared to how frequent it's written today.
Magnus Hagander
issued by autovacuum. Add accessor functions to them, and use those in the
pg_stat_*_tables system views.
Catalog version bumped due to changes in the pgstat views and the pgstat file.
Patch from Larry Rosenman, minor improvements by me.
shutdown, or when requested by a backend:
It changes so the file is only written once every 5 minutes (changeable
of course, I just picked something) instead of once every half second.
It's still written when the stats collector shuts down, just as before.
And it is now also written on backend request. A backend requests a
rewrite by simply sending a special stats message. It operates on the
assumption that the backends aren't actually going to read the
statistics file very often, compared to how frequent it's written today.
Magnus Hagander
Per recent discussion, this seems to be making the stats less accurate
rather than more so, particularly on Windows where PID values may be
reused very quickly. Patch by Peter Brant.
files: avoid creating stats hashtable entries for tables that aren't being
touched except by vacuum/analyze, ensure that entries for dropped tables are
removed promptly, and tweak the data layout to avoid storing useless struct
padding. Also improve the performance of pgstat_vacuum_tabstat(), and make
sure that autovacuum invokes it exactly once per autovac cycle rather than
multiple times or not at all. This should cure recent complaints about 8.1
showing much higher stats I/O volume than was seen in 8.0. It'd still be a
good idea to revisit the design with an eye to not re-writing the entire
stats dataset every half second ... but that would be too much to backpatch,
I fear.
an LWLock instead of a spinlock. This hardly matters on Unix machines
but should improve startup performance on Windows (or any port using
EXEC_BACKEND). Per previous discussion.
file. The original code probed the PGPROC array separately for each PID,
which was not good for large numbers of backends: not only is the runtime
O(N^2) but most of it is spent holding ProcArrayLock. Instead, take the
lock just once and copy the active PIDs into an array, then use qsort
and bsearch so that the lookup time is more like O(N log N).
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.
generated by bitmap index scans. Along the way, simplify and speed up
the code for counting sequential and index scans; it was both confusing
and inefficient to be taking care of that in the per-tuple loops, IMHO.
initdb forced because of internal changes in pg_stat view definitions.
(the stats system has always collected this info, but the views were
filtering it out). Modify autovacuum so that over-threshold activity
in a toast table can trigger a VACUUM of the parent table, even if the
parent didn't appear to need vacuuming itself. Per discussion a month
or so back about "short, wide tables".
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
against the PGPROC array. Anything in the file that isn't in PGPROC
gets rejected as being a stale entry. This should solve complaints about
stale entries in pg_stat_activity after a BETERM message has been dropped
due to overload.
exit, instead of trying to take shortcuts. Introduce some additional
shutdown callback routines to eliminate kluges like having ProcKill
be responsible for shutting down the buffer manager. Ensure that the
order of operations during shutdown is predictable and what you would
expect given the module layering.
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.
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.
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.
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.
collector messages, per recent discussion on pgsql-patches. This
actually required quite a few changes -- for example,
"databaseid != InvalidOid" was used to check whether a slot in the
backend entry table was initialized, but that no longer works since
the slot might be initialized prior to receiving the BESTART message
which contains the database id. We now use procpid > 0 to indicate
that a slot is non-empty.
Other changes:
- various comment improvements and cleanups
- there's no need to zero-out the entire activity buffer in
pgstat_add_backend(), we can just set activity[0] to '\0'.
- remove the counting of the # of connections to a database; this
was not used anywhere
One change in behavior I wasn't sure about: previously, the code
would create a hash table entry for a database as soon as any message
was received whose header referenced that database. Now, we only
create hash table entries as needed (so for example BESTART won't
create a database hash table entry, since it doesn't need to
access anything in the per-db hash table). It would be easy enough
to retain the old behavior, but AFAICS it is not required.
* Add session start time to pg_stat_activity
* Add the client IP address and port to pg_stat_activity
Original patch from Magnus Hagander, code review by Neil Conway. Catalog
version bumped. This patch sends the client IP address and port number in
every statistics message; that's not ideal, but will be fixed up shortly.
whose keys are OIDs. The only one that looks particularly performance
critical is the relcache hashtable, but as long as we've got the function
we may as well use it wherever it's applicable.
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 ...
exit. Without this, operations triggered during backend exit (such as
temp table deletions) won't be counted ... which given heavy usage of
temp tables can lead to pg_autovacuum falling way behind on the need
to vacuum pg_class and pg_attribute. Per reports from Steve Crawford
and others.
even uglier than it was already :-(. Also, on Windows only, use temporary
shared memory segments instead of ordinary files to pass over critical
variable values from postmaster to child processes. Magnus Hagander
returning a NULL pointer (some callers remembered to check the return
value, but some did not -- it is safer to just bail out).
Also, cleanup pgstat.c to use elog(ERROR) rather than elog(LOG) followed
by exit().
recommend that people go get Apache's rotatelogs program. Additional
benefits are that configuration is done through GUC, rather than
externally, and that the postmaster can monitor the log rotator and
restart it after failure (though we certainly hope that won't happen
often).
Andreas Pflug, some rework by Tom Lane.
performance front, but with feature freeze upon us I think it's time to
drive a stake in the ground and say that this will be in 7.5.
Alvaro Herrera, with some help from Tom Lane.
specified in just one place and adhered to exactly, rather than just more
or less. A side effect is to increase PGSTAT_ACTIVITY_SIZE (maximum
reported query length) from 256 to nearly 1000.
begin the shutdown checkpoint; there isn't anything left for them to do,
so we may as well ensure that they shut down sooner rather than later.
Per discussion.
place of time_t, as per prior discussion. The behavior does not change
on machines without a 64-bit-int type, but on machines with one, which
is most, we are rid of the bizarre boundary behavior at the edges of
the 32-bit-time_t range (1901 and 2038). The system will now treat
times over the full supported timestamp range as being in your local
time zone. It may seem a little bizarre to consider that times in
4000 BC are PST or EST, but this is surely at least as reasonable as
propagating Gregorian calendar rules back that far.
I did not modify the format of the zic timezone database files, which
means that for the moment the system will not know about daylight-savings
periods outside the range 1901-2038. Given the way the files are set up,
it's not a simple decision like 'widen to 64 bits'; we have to actually
think about the range of years that need to be supported. We should
probably inquire what the plans of the upstream zic people are before
making any decisions of our own.
than being random pieces of other files. Give bgwriter responsibility
for all checkpoint activity (other than a post-recovery checkpoint);
so this child process absorbs the functionality of the former transient
checkpoint and shutdown subprocesses. While at it, create an actual
include file for postmaster.c, which for some reason never had its own
file before.
about a third, make it work on non-Windows platforms again. (But perhaps
I broke the WIN32 code, since I have no way to test that.) Fold all the
paths that fork postmaster child processes to go through the single
routine SubPostmasterMain, which takes care of resurrecting the state that
would normally be inherited from the postmaster (including GUC variables).
Clean up some places where there's no particularly good reason for the
EXEC and non-EXEC cases to work differently. Take care of one or two
FIXMEs that remained in the code.
* removed a few redundant defines
* get_user_name safe under win32
* rationalized pipe read EOF for win32 (UPDATED PATCH USED)
* changed all backend instances of sleep() to pg_usleep
- except for the SLEEP_ON_ASSERT in assert.c, as it would exceed a
32-bit long [Note to patcher: If a SLEEP_ON_ASSERT of 2000 seconds is
acceptable, please replace with pg_usleep(2000000000L)]
I added a comment to that part of the code:
/*
* It would be nice to use pg_usleep() here, but only does 2000 sec
* or 33 minutes, which seems too short.
*/
sleep(1000000);
Claudio Natoli
It works on the principle of turning sockets into non-blocking, and then
emulate blocking behaviour on top of that, while allowing signals to
run. Signals are now implemented using an event instead of APCs, thus
getting rid of the issue of APCs not being compatible with "old style"
sockets functions.
It also moves the win32 specific code away from pqsignal.h/c into
port/win32, and also removes the "thread style workaround" of the APC
issue previously in place.
In order to make things work, a few things are also changed in pgstat.c:
1) There is now a separate pipe to the collector and the bufferer. This
is required because the pipe will otherwise only be signalled in one of
the processes when the postmaster goes down. The MS winsock code for
select() must have some kind of workaround for this behaviour, but I
have found no stable way of doing that. You really are not supposed to
use the same socket from more than one process (unless you use
WSADuplicateSocket(), in which case the docs specifically say that only
one will be flagged).
2) The check for "postmaster death" is moved into a separate select()
call after the main loop. The previous behaviour select():ed on the
postmaster pipe, while later explicitly saying "we do NOT check for
postmaster exit inside the loop".
The issue was that the code relies on the same select() call seeing both
the postmaster pipe *and* the pgstat pipe go away. This does not always
happen, and it appears that useing WSAEventSelect() makes it even more
common that it does not.
Since it's only called when the process exits, I don't think using a
separate select() call will have any significant impact on how the stats
collector works.
Magnus Hagander
initialization of stats process under EXEC_BACKEND.
[A cleaner, rationalized approach to stat/backend/SSDataBase child
processes under EXEC_BACKEND is on my TODO list. However this patch
takes care of immediate concerns (ie. stats test now passes under
win32)]
Claudio Natoli
whereToSendOutput instead because they are really inquiring about
the correct client communication protocol. Update some comments.
This is pointing towards supporting regular FE/BE client protocol
in a standalone backend, per discussion a month or so back.
PostmasterPid variable, which gets set (early) in PostmasterMain
getppid would not be the postmaster?
[fork/exec] Implements processCancelRequest by keeping an array of
pid/cancel_key structs in shared mem
[fork/exec] Moves AttachSharedMemoryAndSemaphores call for backends into
SubPostmasterMain
[win32] Implements reaper/waitpid by keeping an arrays of children
pids,handles in postmaster local mem
- this item is largely untested, for reasons which should be
obvious, but appears sound
[win32/all] Added extern for pgpipe in Win32 case, and changed the second
pipe call (which seems to have been missed earlier) to pgpipe
[win32] #define'd ftruncate to chsize in the Win32 case
[win32] PG_USLEEP for Win32 has a misplaced paren. Fixed.
[win32] DLLIMPORT handling for MingW case
Claudio Natoli
BackendFork/SSDataBase/pgstat) startup, to allow fork/exec calls to
closely mimic (the soon to be provided) Win32 CreateProcess equivalent
calls.
Claudio Natoli
was modified for IPv6. Use a robust definition of struct sockaddr_storage,
do a proper configure test to see if ss_len exists, don't assume that
getnameinfo() will handle AF_UNIX sockets, don't trust getaddrinfo to
return the protocol we ask for, etc. This incorporates several outstanding
patches from Kurt Roeckx, but I'm to blame for anything that doesn't
work ...
Win32 port is now called 'win32' rather than 'win'
add -lwsock32 on Win32
make gethostname() be only used when kerberos4 is enabled
use /port/getopt.c
new /port/opendir.c routines
disable GUC unix_socket_group on Win32
convert some keywords.c symbols to KEYWORD_P to prevent conflict
create new FCNTL_NONBLOCK macro to turn off socket blocking
create new /include/port.h file that has /port prototypes, move
out of c.h
new /include/port/win32_include dir to hold missing include files
work around ERROR being defined in Win32 includes
I had inadvertently omitted it while rearranging things to support
length-counted incoming messages. Also, change the parser's API back
to accepting a 'char *' query string instead of 'StringInfo', as the
latter wasn't buying us anything except overhead. (I think when I put
it in I had some notion of making the parser API 8-bit-clean, but
seeing that flex depends on null-terminated input, that's not really
ever gonna happen.)
ProcKill instead, where we still have a PGPROC with which to wait on
LWLocks. This fixes 'can't wait without a PROC structure' failures
occasionally seen during backend shutdown (I'm surprised they weren't
more frequent, actually). Add an Assert() to LWLockAcquire to help
catch any similar mistakes in future. Fix failure to update MyProcPid
for standalone backends and pgstat processes.
(overlaying low byte of page size) and add HEAP_HASOID bit to t_infomask,
per earlier discussion. Simplify scheme for overlaying fields in tuple
header (no need for cmax to live in more than one place). Don't try to
clear infomask status bits in tqual.c --- not safe to do it there. Don't
try to force output table of a SELECT INTO to have OIDs, either. Get rid
of unnecessarily complex three-state scheme for TupleDesc.tdhasoids, which
has already caused one recent failure. Improve documentation.
bitmap, if present).
Per Tom Lane's suggestion the information whether a tuple has an oid
or not is carried in the tuple descriptor. For debugging reasons
tdhasoid is of type char, not bool. There are predefined values for
WITHOID, WITHOUTOID and UNDEFOID.
This patch has been generated against a cvs snapshot from last week
and I don't expect it to apply cleanly to current sources. While I
post it here for public review, I'm working on a new version against a
current snapshot. (There's been heavy activity recently; hope to
catch up some day ...)
This is a long patch; if it is too hard to swallow, I can provide it
in smaller pieces:
Part 1: Accessor macros
Part 2: tdhasoid in TupDesc
Part 3: Regression test
Part 4: Parameter withoid to heap_addheader
Part 5: Eliminate t_oid from HeapTupleHeader
Part 2 is the most hairy part because of changes in the executor and
even in the parser; the other parts are straightforward.
Up to part 4 the patched postmaster stays binary compatible to
databases created with an unpatched version. Part 5 is small (100
lines) and finally breaks compatibility.
Manfred Koizar
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.
As proof of concept, provide an alternate implementation based on POSIX
semaphores. Also push the SysV shared-memory implementation into a
separate file so that it can be replaced conveniently.
o Change all current CVS messages of NOTICE to WARNING. We were going
to do this just before 7.3 beta but it has to be done now, as you will
see below.
o Change current INFO messages that should be controlled by
client_min_messages to NOTICE.
o Force remaining INFO messages, like from EXPLAIN, VACUUM VERBOSE, etc.
to always go to the client.
o Remove INFO from the client_min_messages options and add NOTICE.
Seems we do need three non-ERROR elog levels to handle the various
behaviors we need for these messages.
Regression passed.
environment strings need to be moved around, do so when called from
initial startup (main.c), not in init_ps_status. This eliminates the
former risk of invalidating saved environment-string pointers, since
no code has yet had a chance to grab any such pointers when main.c
is running.
per suggestion from Peter. Simplify several APIs by transmitting the
original argv location directly from main.c to ps_status.c, instead of
passing it down through several levels of subroutines.
readability. Bizarre '(long *) TRUE' return convention is gone,
in favor of just raising an error internally in dynahash.c when
we detect hashtable corruption. HashTableWalk is gone, in favor
of using hash_seq_search directly, since it had no hope of working
with non-LONGALIGNable datatypes. Simplify some other code that was
made undesirably grotty by promixity to HashTableWalk.
Partial support for BEOS (not sure whether second fork of grandchild
process needs these extra calls or not; someone who has BEOS will need
to test it).
portability issues). Caller-visible data structures are now allocated
on MAXALIGN boundaries, allowing safe use of datatypes wider than 'long'.
Rejigger hash_create API so that caller specifies size of key and
total size of entry, not size of key and size of rest of entry.
This simplifies life considerably since each number is just a sizeof(),
and padding issues etc. are taken care of automatically.
Make sure it exits immediately when collector process dies --- in old code,
buffer process would hang around and compete with the new buffer process
for packets. Make sure it doesn't block on writing the pipe when the
collector falls more than a pipeload behind. Avoid leaking pgstats FDs
into every backend.
platforms system(2) gets confused unless the signal handler is set to
SIG_DFL, not SIG_IGN. pgstats.c now uses pqsignal() as it should,
not signal(). Also, arrange for the stats collector process to show
a reasonable ID in 'ps', rather than looking like a postmaster.
Tom Lane
Magnus Hagander
Alvaro Herrera
Bruce Momjian
Neil Conway
Peter Eisentraut
PostgreSQL Daemon
Tatsuo Ishii
Jan Wieck