public inbox for [email protected]
help / color / mirror / Atom feedFrom: Joel Jacobson <[email protected]>
To: pgsql-hackers <[email protected]>
Subject: Re: Optimize LISTEN/NOTIFY
Date: Sat, 08 Nov 2025 16:04:42 +0100
Message-ID: <[email protected]> (raw)
In-Reply-To: <[email protected]>
References: <[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<CAK80=jhmE40KVqQ3ho37MArS7cAED1p9m7uikDxcnDmqdW7t8A@mail.gmail.com>
<[email protected]>
<[email protected]>
<CA+hUKGLrMGkWDB0cwTa0RqD+AF7O-Ywgck8aVYKwOQnZgYRRug@mail.gmail.com>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<CAFY6G8dap-bCnAnMG-2Gzew8yv2Vbi9gsx9+yszKMmd57ygfvA@mail.gmail.com>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<CAE7r3MLivh1sHWF06hrVXkiQbw-KChPcQsh+9CheXprm5vRVMQ@mail.gmail.com>
<[email protected]>
<CAE7r3MK-3AOdh1mpZ8hw9h6F_i0D5RMoAy7CttnfCJRpB8GJDA@mail.gmail.com>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
<CAE7r3MJxxycYwDFJdTJ+HmzeNJ_JGKyr4HpT5njN1ouVX56OUg@mail.gmail.com>
<[email protected]>
<CAE7r3MJ0VoAJzdLzX0dgfPJpBJxW+wg_pYaCi6mQJi47+qukhg@mail.gmail.com>
<[email protected]>
<[email protected]>
<[email protected]>
<[email protected]>
On Sat, Nov 8, 2025, at 13:59, Joel Jacobson wrote:
> On Fri, Nov 7, 2025, at 19:59, Joel Jacobson wrote:
>> * The logic in SignalBackends has been reworked and simplified,
>> thanks to the new isAdvancing and advancingPos fields.
>> I now think it's finally easy to reason about why each branch
>> in SignalBackends must be correct.
>
> I was wrong. I wrongly assumed asyncQueueReadAllNotifications would read
> up until head, which it might not actually do:
>
> * Process messages up to the stop position, end of page, or an
> * uncommitted message.
>
> This in turn could cause a listening backend to remain behind, if there
> would be no more notifies, so it unfortunately seems like we will always
> need to signal when a backend isAdvancing, and therefore have no use of
> the advancingPos field.
>
> I will do more correctness and benchmark testing before posting a new
> version. Just wanted to give you a heads up on the bug, so you don't
> waste time reviewing.
Changes since v24:
* Removed the QueuePosition advancingPos QueueBackendStatus field.
* Always signal when isAdvancing is true.
Some benchmarking:
master:
% ./pg_async_notify_test --listeners 1 --notifiers 1 --channels 1000 --sleep 0.1 --sleep-exp 1.01
13 s: 8668 sent (658/s), 8676 received (653/s)
Notification Latency Distribution:
0.00-0.01ms 0 (0.0%) avg: 0.000ms
0.01-0.10ms 0 (0.0%) avg: 0.000ms
0.10-1.00ms # 3 (0.0%) avg: 0.783ms
1.00-10.00ms # 49 (0.6%) avg: 5.559ms
10.00-100.00ms # 168 (1.9%) avg: 56.360mss
>100.00ms ######### 8456 (97.5%) avg: 256.086ms
v25:
% ./pg_async_notify_test --listeners 1 --notifiers 1 --channels 1000 --sleep 0.1 --sleep-exp 1.01
14 s: 27097 sent (1959/s), 27097 received (1960/s)
Notification Latency Distribution:
0.00-0.01ms 0 (0.0%) avg: 0.000ms
0.01-0.10ms # 962 (3.6%) avg: 0.081ms
0.10-1.00ms ######### 25066 (92.5%) avg: 0.321ms
1.00-10.00ms # 1069 (3.9%) avg: 2.104ms
10.00-100.00ms 0 (0.0%) avg: 0.000ms
>100.00ms 0 (0.0%) avg: 0.000ms
On master, I see lots of "waiting for AccessExclusiveLock on object 0"
in the logs for this benchmark setup, but none at all for v25.
I wonder if there would be some way to solve the problem with v24,
without having to always signal when isAdvancing is true with an
advancingPos at or ahead of the notifier's queueHeadAfterWrite. One idea
I had, that I think is a bad idea, but just mentioning it in case you
think it could work, is to check if the new pos is at head, in
asyncQueueReadAllNotifications's PG_FINALLY block, and if not, to let
the listening backend signal itself. I guess this is a bad idea, since
then it could signal itself over and over again, and consume a lot of
resources, if some uncommitted message takes a long time to commit. The
current mechanism limit the rate of wake-ups to the same rate as the
notifies, which seems sensible. Thoughts on this?
/Joel
Attachments:
[application/octet-stream] 0001-optimize_listen_notify-v25.patch (9.3K, 2-0001-optimize_listen_notify-v25.patch)
download | inline diff:
From def00569ec6d23b3652edf6a3727e007899fafe8 Mon Sep 17 00:00:00 2001
From: Joel Jacobson <[email protected]>
Date: Wed, 8 Oct 2025 09:30:54 +0200
Subject: [PATCH 1/2] Improve LISTEN/NOTIFY test coverage
This adds isolation tests to cover previously untested code paths:
* Check simple NOTIFY reparenting when parent has no action
* Check LISTEN reparenting in subtransaction
* Check LISTEN merge path when both outer and inner transactions have actions
* Check LISTEN abort path (ROLLBACK TO SAVEPOINT discards pending actions)
* Check notification_match function (triggered by hash table duplicate detection)
* Check that notifications sent from a backend that has not done LISTEN
are properly delivered to a listener in another backend
This also adds a test to prepare for the next patch:
* Check ChannelHashAddListener array growth
---
src/test/isolation/expected/async-notify.out | 114 ++++++++++++++++++-
src/test/isolation/specs/async-notify.spec | 68 +++++++++++
2 files changed, 181 insertions(+), 1 deletion(-)
diff --git a/src/test/isolation/expected/async-notify.out b/src/test/isolation/expected/async-notify.out
index 556e1805893..443a6eb669f 100644
--- a/src/test/isolation/expected/async-notify.out
+++ b/src/test/isolation/expected/async-notify.out
@@ -1,4 +1,4 @@
-Parsed test spec with 3 sessions
+Parsed test spec with 7 sessions
starting permutation: listenc notify1 notify2 notify3 notifyf
step listenc: LISTEN c1; LISTEN c2;
@@ -47,6 +47,105 @@ notifier: NOTIFY "c2" with payload "payload" from notifier
notifier: NOTIFY "c1" with payload "payloads" from notifier
notifier: NOTIFY "c2" with payload "payloads" from notifier
+starting permutation: listenc notifys_simple
+step listenc: LISTEN c1; LISTEN c2;
+step notifys_simple:
+ BEGIN;
+ SAVEPOINT s1;
+ NOTIFY c1, 'simple1';
+ NOTIFY c2, 'simple2';
+ RELEASE SAVEPOINT s1;
+ COMMIT;
+
+notifier: NOTIFY "c1" with payload "simple1" from notifier
+notifier: NOTIFY "c2" with payload "simple2" from notifier
+
+starting permutation: lsbegin lssavepoint lslisten lsrelease lscommit lsnotify
+step lsbegin: BEGIN;
+step lssavepoint: SAVEPOINT s1;
+step lslisten: LISTEN c1; LISTEN c2;
+step lsrelease: RELEASE SAVEPOINT s1;
+step lscommit: COMMIT;
+step lsnotify: NOTIFY c1, 'subxact_test';
+listen_subxact: NOTIFY "c1" with payload "subxact_test" from listen_subxact
+
+starting permutation: lsbegin lslisten_outer lssavepoint lslisten lsrelease lscommit lsnotify
+step lsbegin: BEGIN;
+step lslisten_outer: LISTEN c3;
+step lssavepoint: SAVEPOINT s1;
+step lslisten: LISTEN c1; LISTEN c2;
+step lsrelease: RELEASE SAVEPOINT s1;
+step lscommit: COMMIT;
+step lsnotify: NOTIFY c1, 'subxact_test';
+listen_subxact: NOTIFY "c1" with payload "subxact_test" from listen_subxact
+
+starting permutation: lsbegin lssavepoint lslisten lsrollback lscommit lsnotify_check
+step lsbegin: BEGIN;
+step lssavepoint: SAVEPOINT s1;
+step lslisten: LISTEN c1; LISTEN c2;
+step lsrollback: ROLLBACK TO SAVEPOINT s1;
+step lscommit: COMMIT;
+step lsnotify_check: NOTIFY c1, 'should_not_receive';
+
+starting permutation: listenc notify_many_with_dup
+step listenc: LISTEN c1; LISTEN c2;
+step notify_many_with_dup:
+ BEGIN;
+ SELECT pg_notify('c1', 'msg' || s::text) FROM generate_series(1, 17) s;
+ SELECT pg_notify('c1', 'msg1');
+ COMMIT;
+
+pg_notify
+---------
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+(17 rows)
+
+pg_notify
+---------
+
+(1 row)
+
+notifier: NOTIFY "c1" with payload "msg1" from notifier
+notifier: NOTIFY "c1" with payload "msg2" from notifier
+notifier: NOTIFY "c1" with payload "msg3" from notifier
+notifier: NOTIFY "c1" with payload "msg4" from notifier
+notifier: NOTIFY "c1" with payload "msg5" from notifier
+notifier: NOTIFY "c1" with payload "msg6" from notifier
+notifier: NOTIFY "c1" with payload "msg7" from notifier
+notifier: NOTIFY "c1" with payload "msg8" from notifier
+notifier: NOTIFY "c1" with payload "msg9" from notifier
+notifier: NOTIFY "c1" with payload "msg10" from notifier
+notifier: NOTIFY "c1" with payload "msg11" from notifier
+notifier: NOTIFY "c1" with payload "msg12" from notifier
+notifier: NOTIFY "c1" with payload "msg13" from notifier
+notifier: NOTIFY "c1" with payload "msg14" from notifier
+notifier: NOTIFY "c1" with payload "msg15" from notifier
+notifier: NOTIFY "c1" with payload "msg16" from notifier
+notifier: NOTIFY "c1" with payload "msg17" from notifier
+
+starting permutation: listenc llisten l2listen l3listen lslisten
+step listenc: LISTEN c1; LISTEN c2;
+step llisten: LISTEN c1; LISTEN c2;
+step l2listen: LISTEN c1;
+step l3listen: LISTEN c1;
+step lslisten: LISTEN c1; LISTEN c2;
+
starting permutation: llisten notify1 notify2 notify3 notifyf lcheck
step llisten: LISTEN c1; LISTEN c2;
step notify1: NOTIFY c1;
@@ -95,6 +194,8 @@ listener: NOTIFY "c2" with payload "" from notifier
starting permutation: l2listen l2begin notify1 lbegins llisten lcommit l2commit l2stop
step l2listen: LISTEN c1;
+listener2: NOTIFY "c1" with payload "" from notifier
+listener2: NOTIFY "c1" with payload "" from notifier
step l2begin: BEGIN;
step notify1: NOTIFY c1;
step lbegins: BEGIN ISOLATION LEVEL SERIALIZABLE;
@@ -104,6 +205,17 @@ step l2commit: COMMIT;
listener2: NOTIFY "c1" with payload "" from notifier
step l2stop: UNLISTEN *;
+starting permutation: lch_listen nch_notify lch_check
+step lch_listen: LISTEN ch;
+step nch_notify: NOTIFY ch, 'aa';
+step lch_check: SELECT 1 AS x;
+x
+-
+1
+(1 row)
+
+listener_ch: NOTIFY "ch" with payload "aa" from notifier_ch
+
starting permutation: llisten lbegin usage bignotify usage
step llisten: LISTEN c1; LISTEN c2;
step lbegin: BEGIN;
diff --git a/src/test/isolation/specs/async-notify.spec b/src/test/isolation/specs/async-notify.spec
index 0b8cfd91083..0a01e777b98 100644
--- a/src/test/isolation/specs/async-notify.spec
+++ b/src/test/isolation/specs/async-notify.spec
@@ -31,6 +31,20 @@ step notifys1 {
ROLLBACK TO SAVEPOINT s2;
COMMIT;
}
+step notifys_simple {
+ BEGIN;
+ SAVEPOINT s1;
+ NOTIFY c1, 'simple1';
+ NOTIFY c2, 'simple2';
+ RELEASE SAVEPOINT s1;
+ COMMIT;
+}
+step notify_many_with_dup {
+ BEGIN;
+ SELECT pg_notify('c1', 'msg' || s::text) FROM generate_series(1, 17) s;
+ SELECT pg_notify('c1', 'msg1');
+ COMMIT;
+}
step usage { SELECT pg_notification_queue_usage() > 0 AS nonzero; }
step bignotify { SELECT count(pg_notify('c1', s::text)) FROM generate_series(1, 1000) s; }
teardown { UNLISTEN *; }
@@ -53,6 +67,38 @@ step l2begin { BEGIN; }
step l2commit { COMMIT; }
step l2stop { UNLISTEN *; }
+# Third listener session for testing array growth.
+
+session listener3
+step l3listen { LISTEN c1; }
+teardown { UNLISTEN *; }
+
+# Listener session for cross-session notification test with channel 'ch'.
+
+session listener_ch
+step lch_listen { LISTEN ch; }
+step lch_check { SELECT 1 AS x; }
+teardown { UNLISTEN *; }
+
+# Notifier session for cross-session notification test with channel 'ch'.
+
+session notifier_ch
+step nch_notify { NOTIFY ch, 'aa'; }
+
+# Session for testing LISTEN in subtransaction with separate steps.
+
+session listen_subxact
+step lsbegin { BEGIN; }
+step lslisten_outer { LISTEN c3; }
+step lssavepoint { SAVEPOINT s1; }
+step lslisten { LISTEN c1; LISTEN c2; }
+step lsrelease { RELEASE SAVEPOINT s1; }
+step lsrollback { ROLLBACK TO SAVEPOINT s1; }
+step lscommit { COMMIT; }
+step lsnotify { NOTIFY c1, 'subxact_test'; }
+step lsnotify_check { NOTIFY c1, 'should_not_receive'; }
+teardown { UNLISTEN *; }
+
# Trivial cases.
permutation listenc notify1 notify2 notify3 notifyf
@@ -60,6 +106,24 @@ permutation listenc notify1 notify2 notify3 notifyf
# Check simple and less-simple deduplication.
permutation listenc notifyd1 notifyd2 notifys1
+# Check simple NOTIFY reparenting when parent has no action.
+permutation listenc notifys_simple
+
+# Check LISTEN reparenting in subtransaction.
+permutation lsbegin lssavepoint lslisten lsrelease lscommit lsnotify
+
+# Check LISTEN merge path when both outer and inner transactions have actions.
+permutation lsbegin lslisten_outer lssavepoint lslisten lsrelease lscommit lsnotify
+
+# Check LISTEN abort path (ROLLBACK TO SAVEPOINT discards pending actions).
+permutation lsbegin lssavepoint lslisten lsrollback lscommit lsnotify_check
+
+# Check notification_match function (triggered by hash table duplicate detection).
+permutation listenc notify_many_with_dup
+
+# Check ChannelHashAddListener array growth.
+permutation listenc llisten l2listen l3listen lslisten
+
# Cross-backend notification delivery. We use a "select 1" to force the
# listener session to check for notifies. In principle we could just wait
# for delivery, but that would require extra support in isolationtester
@@ -73,6 +137,10 @@ permutation listenc llisten notify1 notify2 notify3 notifyf lcheck
# and notify queue is not empty
permutation l2listen l2begin notify1 lbegins llisten lcommit l2commit l2stop
+# Check that notifications sent from a backend that has not done LISTEN
+# are properly delivered to a listener in another backend.
+permutation lch_listen nch_notify lch_check
+
# Verify that pg_notification_queue_usage correctly reports a non-zero result,
# after submitting notifications while another connection is listening for
# those notifications and waiting inside an active transaction. We have to
--
2.50.1
[application/octet-stream] 0002-optimize_listen_notify-v25.patch (42.7K, 3-0002-optimize_listen_notify-v25.patch)
download | inline diff:
From 2abed1a217642a9037fe03b270906c849662c3a9 Mon Sep 17 00:00:00 2001
From: Joel Jacobson <[email protected]>
Date: Sat, 8 Nov 2025 13:47:09 +0100
Subject: [PATCH 2/2] Optimize LISTEN/NOTIFY with shared channel map and direct
advancement
This patch reworks the LISTEN/NOTIFY signaling path to avoid the
long-standing inefficiency where every commit wakes all listening
backends in the same database, even those that are listening on
completely different channels.
Problem
-------
At present, SignalBackends has no central knowledge of which backend
listens on which channel. When a backend commits a transaction that
issued NOTIFY, it simply iterates over all registered listeners in the
same database and sends each one a PROCSIG_NOTIFY_INTERRUPT signal.
That behavior is fine when all listeners are on the same channel, but
when many backends are listening on different channels, each NOTIFY
triggers a storm of unnecessary wakeups and context switches. As the
number of idle listeners grows, this often becomes the bottleneck and
throughput drops sharply.
Overview of the solution
------------------------
This patch introduces a lazily-created dynamic shared hash (dshash)
backed by dynamic shared memory (DSA) that maps (dboid, channel) to
arrays of listening backends (ProcNumbers). This allows the sender to
target only those backends actually listening on the channels for which
it has queued notifications.
At commit time:
* AtCommit_Notify updates the shared channelHash to reflect any LISTEN
or UNLISTEN actions performed in the transaction.
* SignalBackends consults this hash to find the backends that are
listening on the channels being notified in the current database, and
signals only those.
Each backend's entry in AsyncQueueControl now includes a wakeupPending
flag to prevent duplicate signals while a previous wakeup is still being
processed.
Direct advancement
------------------
A further optimization avoids signaling idle backends that are not
listening on any of the channels notified within the transaction.
While queuing notifications, PreCommit_Notify records the queue head
position both before and after writing its notifications. Because all
writers are serialized by the existing cluster-wide heavyweight lock on
"database 0", no backend (from any database) can insert entries between
those two points. This guarantees that the region [oldHead, newHead)
contains only the entries written by our commit.
SignalBackends uses this fact to directly advance any backend still
positioned at oldHead up to newHead, avoiding a needless wakeup for
listeners that would otherwise not find any notifies of interest.
To handle advancing backends correctly, each backend's entry tracks both
whether it is currently advancing (isAdvancing) and the target position
it is advancing to (advancingPos). This allows SignalBackends to signal
advancing backends only when their target position would leave them
behind the new queue head, while safely direct-advancing idle backends
that would not be interested in the newly written notifications.
Other notes
-----------
* Maintains dual data structures: a shared channelHash for determining
which backends to signal, and a local per-backend listenChannels list
for fast lock-free lookups during notification processing. This avoids
contention on the shared hash during the high-frequency IsListeningOn
checks that occur for every notification read from the queue.
* Backends remain registered in the global listener list as long as
listenChannels is non-empty.
* Adds LWLock tranche NOTIFY_CHANNEL_HASH and wait event
NotifyChannelHash for visibility.
* No user-visible behavioral changes; this is an internal optimization
only.
---
src/backend/commands/async.c | 708 +++++++++++++++---
.../utils/activity/wait_event_names.txt | 1 +
src/include/storage/lwlocklist.h | 1 +
src/tools/pgindent/typedefs.list | 2 +
4 files changed, 608 insertions(+), 104 deletions(-)
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4bd37d5beb5..1f54fa8e41b 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -24,8 +24,10 @@
* All notification messages are placed in the queue and later read out
* by listening backends.
*
- * There is no central knowledge of which backend listens on which channel;
- * every backend has its own list of interesting channels.
+ * We also maintain a dynamic shared hash table (dshash) that maps channel
+ * names to the set of backends listening on each channel. This table is
+ * created lazily on the first LISTEN command and grows dynamically as
+ * needed.
*
* Although there is only one queue, notifications are treated as being
* database-local; this is done by including the sender's database OID
@@ -68,16 +70,21 @@
* CommitTransaction() which will then do the actual transaction commit.
*
* After commit we are called another time (AtCommit_Notify()). Here we
- * make any actual updates to the effective listen state (listenChannels).
- * Then we signal any backends that may be interested in our messages
- * (including our own backend, if listening). This is done by
- * SignalBackends(), which scans the list of listening backends and sends a
- * PROCSIG_NOTIFY_INTERRUPT signal to every listening backend (we don't
- * know which backend is listening on which channel so we must signal them
- * all). We can exclude backends that are already up to date, though, and
- * we can also exclude backends that are in other databases (unless they
- * are way behind and should be kicked to make them advance their
- * pointers).
+ * make any actual updates to the local listen state (listenChannelsHash) and
+ * shared channel hash table (channelHash). Then we signal any backends
+ * that may be interested in our messages (including our own backend,
+ * if listening). This is done by SignalBackends(), which consults the
+ * shared channel hash table to identify listeners for the channels that
+ * have pending notifications in the current database. Each selected
+ * backend is marked as having a wakeup pending to avoid duplicate signals,
+ * and a PROCSIG_NOTIFY_INTERRUPT signal is sent to it.
+ *
+ * When writing notifications, PreCommit_Notify() records the queue head
+ * position both before and after the write. Because all writers serialize
+ * on a cluster-wide heavyweight lock, no backend can insert entries between
+ * these two points. SignalBackends() uses this fact to directly advance any
+ * backend that is still positioned at the old head, avoiding unnecessary
+ * wakeups for idle listeners that have nothing to read.
*
* Finally, after we are out of the transaction altogether and about to go
* idle, we scan the queue for messages that need to be sent to our
@@ -137,14 +144,17 @@
#include "commands/async.h"
#include "common/hashfn.h"
#include "funcapi.h"
+#include "lib/dshash.h"
#include "libpq/libpq.h"
#include "libpq/pqformat.h"
#include "miscadmin.h"
+#include "storage/dsm_registry.h"
#include "storage/ipc.h"
#include "storage/lmgr.h"
#include "storage/procsignal.h"
#include "tcop/tcopprot.h"
#include "utils/builtins.h"
+#include "utils/dsa.h"
#include "utils/guc_hooks.h"
#include "utils/memutils.h"
#include "utils/ps_status.h"
@@ -162,6 +172,29 @@
*/
#define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
+/*
+ * Channel hash table definitions
+ *
+ * This hash table maps (database OID, channel name) keys to arrays of
+ * ProcNumbers representing the backends listening on each channel.
+ */
+
+#define INITIAL_LISTENERS_ARRAY_SIZE 4
+
+typedef struct ChannelHashKey
+{
+ Oid dboid;
+ char channel[NAMEDATALEN];
+} ChannelHashKey;
+
+typedef struct ChannelEntry
+{
+ ChannelHashKey key;
+ dsa_pointer listenersArray; /* DSA pointer to ProcNumber array */
+ int numListeners; /* Number of listeners currently stored */
+ int allocatedListeners; /* Allocated size of array */
+} ChannelEntry;
+
/*
* Struct representing an entry in the global notify queue
*
@@ -224,11 +257,16 @@ typedef struct QueuePosition
(x).page != (y).page ? (x) : \
(x).offset > (y).offset ? (x) : (y))
+/* returns true if x comes before y in queue order */
+#define QUEUE_POS_PRECEDES(x,y) \
+ (asyncQueuePagePrecedes((x).page, (y).page) || \
+ ((x).page == (y).page && (x).offset < (y).offset))
+
/*
* Parameter determining how often we try to advance the tail pointer:
* we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
- * also the distance by which a backend in another database needs to be
- * behind before we'll decide we need to wake it up to advance its pointer.
+ * also the distance by which a backend needs to be behind before we'll
+ * decide we need to wake it up to advance its pointer.
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
@@ -246,6 +284,8 @@ typedef struct QueueBackendStatus
Oid dboid; /* backend's database OID, or InvalidOid */
ProcNumber nextListener; /* id of next listener, or INVALID_PROC_NUMBER */
QueuePosition pos; /* backend has read queue up to here */
+ bool wakeupPending; /* signal sent but not yet processed */
+ bool isAdvancing; /* backend is advancing its position */
} QueueBackendStatus;
/*
@@ -260,9 +300,10 @@ typedef struct QueueBackendStatus
* (since no other backend will inspect it).
*
* When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
- * entries of other backends and also change the head pointer. When holding
- * both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE mode, backends
- * can change the tail pointers.
+ * entries of other backends and also change the head pointer. They can
+ * also advance other backends' queue positions, unless they are not
+ * in the process of doing that themselves. When holding both NotifyQueueLock and
+ * NotifyQueueTailLock in EXCLUSIVE mode, backends can change the tail pointers.
*
* SLRU buffer pool is divided in banks and bank wise SLRU lock is used as
* the control lock for the pg_notify SLRU buffers.
@@ -288,11 +329,16 @@ typedef struct AsyncQueueControl
ProcNumber firstListener; /* id of first listener, or
* INVALID_PROC_NUMBER */
TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
+ dsa_handle channelHashDSA;
+ dshash_table_handle channelHashDSH;
QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
} AsyncQueueControl;
static AsyncQueueControl *asyncQueueControl;
+static dsa_area *channelDSA = NULL;
+static dshash_table *channelHash = NULL;
+
#define QUEUE_HEAD (asyncQueueControl->head)
#define QUEUE_TAIL (asyncQueueControl->tail)
#define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
@@ -301,6 +347,8 @@ static AsyncQueueControl *asyncQueueControl;
#define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
#define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
#define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
+#define QUEUE_BACKEND_WAKEUP_PENDING(i) (asyncQueueControl->backend[i].wakeupPending)
+#define QUEUE_BACKEND_IS_ADVANCING(i) (asyncQueueControl->backend[i].isAdvancing)
/*
* The SLRU buffer area through which we access the notification queue
@@ -313,16 +361,16 @@ static SlruCtlData NotifyCtlData;
#define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
/*
- * listenChannels identifies the channels we are actually listening to
- * (ie, have committed a LISTEN on). It is a simple list of channel names,
+ * listenChannelsHash identifies the channels we are actually listening to
+ * (ie, have committed a LISTEN on). It is a hash table of channel names,
* allocated in TopMemoryContext.
*/
-static List *listenChannels = NIL; /* list of C strings */
+static HTAB *listenChannelsHash = NULL;
/*
* State for pending LISTEN/UNLISTEN actions consists of an ordered list of
* all actions requested in the current transaction. As explained above,
- * we don't actually change listenChannels until we reach transaction commit.
+ * we don't actually change listenChannelsHash until we reach transaction commit.
*
* The list is kept in CurTransactionContext. In subtransactions, each
* subtransaction has its own list in its own CurTransactionContext, but
@@ -391,6 +439,7 @@ typedef struct NotificationList
int nestingLevel; /* current transaction nesting depth */
List *events; /* list of Notification structs */
HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
+ HTAB *channelHashtab; /* hash of unique channel names, or NULL */
struct NotificationList *upper; /* details for upper transaction levels */
} NotificationList;
@@ -401,6 +450,11 @@ struct NotificationHash
Notification *event; /* => the actual Notification struct */
};
+struct ChannelHash
+{
+ char channel[NAMEDATALEN];
+};
+
static NotificationList *pendingNotifies = NULL;
/*
@@ -418,6 +472,20 @@ static bool unlistenExitRegistered = false;
/* True if we're currently registered as a listener in asyncQueueControl */
static bool amRegisteredListener = false;
+/*
+ * Queue head positions for direct advancement.
+ * These are captured during PreCommit_Notify while holding the heavyweight
+ * lock on database 0, ensuring no other backend can insert notifications
+ * between them. SignalBackends uses these to advance idle backends.
+ */
+static QueuePosition queueHeadBeforeWrite;
+static QueuePosition queueHeadAfterWrite;
+
+/*
+ * List of channels with pending notifications in the current transaction.
+ */
+static List *pendingNotifyChannels = NIL;
+
/* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
static bool tryAdvanceTail = false;
@@ -428,7 +496,6 @@ bool Trace_notify = false;
int max_notify_queue_pages = 1048576;
/* local function prototypes */
-static inline int64 asyncQueuePageDiff(int64 p, int64 q);
static inline bool asyncQueuePagePrecedes(int64 p, int64 q);
static void queue_listen(ListenActionKind action, const char *channel);
static void Async_UnlistenOnExit(int code, Datum arg);
@@ -457,16 +524,9 @@ static void AddEventToPendingNotifies(Notification *n);
static uint32 notification_hash(const void *key, Size keysize);
static int notification_match(const void *key1, const void *key2, Size keysize);
static void ClearPendingActionsAndNotifies(void);
-
-/*
- * Compute the difference between two queue page numbers.
- * Previously this function accounted for a wraparound.
- */
-static inline int64
-asyncQueuePageDiff(int64 p, int64 q)
-{
- return p - q;
-}
+static inline void ChannelHashPrepareKey(ChannelHashKey *key, Oid dboid, const char *channel);
+static dshash_hash channelHashFunc(const void *key, size_t size, void *arg);
+static void initChannelHash(void);
/*
* Determines whether p precedes q.
@@ -478,6 +538,105 @@ asyncQueuePagePrecedes(int64 p, int64 q)
return p < q;
}
+/*
+ * channelHashFunc
+ * Hash function for channel keys.
+ */
+static dshash_hash
+channelHashFunc(const void *key, size_t size, void *arg)
+{
+ const ChannelHashKey *k = (const ChannelHashKey *) key;
+ dshash_hash h;
+
+ h = DatumGetUInt32(hash_uint32(k->dboid));
+ h ^= DatumGetUInt32(hash_any((const unsigned char *) k->channel,
+ strnlen(k->channel, NAMEDATALEN)));
+
+ return h;
+}
+
+/* parameters for the channel hash table */
+static const dshash_parameters channelDSHParams = {
+ sizeof(ChannelHashKey),
+ sizeof(ChannelEntry),
+ dshash_memcmp,
+ channelHashFunc,
+ dshash_memcpy,
+ LWTRANCHE_NOTIFY_CHANNEL_HASH
+};
+
+/*
+ * initChannelHash
+ * Lazy initialization of the channel hash table.
+ */
+static void
+initChannelHash(void)
+{
+ MemoryContext oldcontext;
+
+ /* Quick exit if we already did this */
+ if (asyncQueueControl->channelHashDSH != DSHASH_HANDLE_INVALID &&
+ channelHash != NULL)
+ return;
+
+ /* Otherwise, use a lock to ensure only one process creates the table */
+ LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
+
+ /* Be sure any local memory allocated by DSA routines is persistent */
+ oldcontext = MemoryContextSwitchTo(TopMemoryContext);
+
+ if (asyncQueueControl->channelHashDSH == DSHASH_HANDLE_INVALID)
+ {
+ /* Initialize dynamic shared hash table for channel hash */
+ channelDSA = dsa_create(LWTRANCHE_NOTIFY_CHANNEL_HASH);
+ dsa_pin(channelDSA);
+ dsa_pin_mapping(channelDSA);
+ channelHash = dshash_create(channelDSA, &channelDSHParams, NULL);
+
+ /* Store handles in shared memory for other backends to use */
+ asyncQueueControl->channelHashDSA = dsa_get_handle(channelDSA);
+ asyncQueueControl->channelHashDSH =
+ dshash_get_hash_table_handle(channelHash);
+ }
+ else if (!channelHash)
+ {
+ /* Attach to existing dynamic shared hash table */
+ channelDSA = dsa_attach(asyncQueueControl->channelHashDSA);
+ dsa_pin_mapping(channelDSA);
+ channelHash = dshash_attach(channelDSA, &channelDSHParams,
+ asyncQueueControl->channelHashDSH,
+ NULL);
+ }
+
+ MemoryContextSwitchTo(oldcontext);
+ LWLockRelease(NotifyQueueLock);
+}
+
+/*
+ * initListenChannelsHash
+ * Lazy initialization of the local listen channels hash table.
+ */
+static void
+initListenChannelsHash(void)
+{
+ HASHCTL hash_ctl;
+
+ /* Quick exit if we already did this */
+ if (listenChannelsHash != NULL)
+ return;
+
+ /* Initialize local hash table for this backend's listened channels */
+ memset(&hash_ctl, 0, sizeof(hash_ctl));
+ hash_ctl.keysize = NAMEDATALEN;
+ hash_ctl.entrysize = sizeof(struct ChannelHash);
+
+ listenChannelsHash =
+ hash_create("Listen Channels",
+ 64,
+ &hash_ctl,
+ HASH_ELEM | HASH_STRINGS);
+}
+
/*
* Report space needed for our shared memory area
*/
@@ -521,12 +680,16 @@ AsyncShmemInit(void)
QUEUE_STOP_PAGE = 0;
QUEUE_FIRST_LISTENER = INVALID_PROC_NUMBER;
asyncQueueControl->lastQueueFillWarn = 0;
+ asyncQueueControl->channelHashDSA = DSA_HANDLE_INVALID;
+ asyncQueueControl->channelHashDSH = DSHASH_HANDLE_INVALID;
for (int i = 0; i < MaxBackends; i++)
{
QUEUE_BACKEND_PID(i) = InvalidPid;
QUEUE_BACKEND_DBOID(i) = InvalidOid;
QUEUE_NEXT_LISTENER(i) = INVALID_PROC_NUMBER;
SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
+ QUEUE_BACKEND_WAKEUP_PENDING(i) = false;
+ QUEUE_BACKEND_IS_ADVANCING(i) = false;
}
}
@@ -657,6 +820,7 @@ Async_Notify(const char *channel, const char *payload)
notifies->events = list_make1(n);
/* We certainly don't need a hashtable yet */
notifies->hashtab = NULL;
+ notifies->channelHashtab = NULL;
notifies->upper = pendingNotifies;
pendingNotifies = notifies;
}
@@ -683,7 +847,7 @@ Async_Notify(const char *channel, const char *payload)
* Common code for listen, unlisten, unlisten all commands.
*
* Adds the request to the list of pending actions.
- * Actual update of the listenChannels list happens during transaction
+ * Actual update of the listenChannelsHash happens during transaction
* commit.
*/
static void
@@ -783,30 +947,49 @@ Async_UnlistenAll(void)
* SQL function: return a set of the channel names this backend is actively
* listening to.
*
- * Note: this coding relies on the fact that the listenChannels list cannot
+ * Note: this coding relies on the fact that the listenChannelsHash cannot
* change within a transaction.
*/
Datum
pg_listening_channels(PG_FUNCTION_ARGS)
{
FuncCallContext *funcctx;
+ HASH_SEQ_STATUS *status;
/* stuff done only on the first call of the function */
if (SRF_IS_FIRSTCALL())
{
+ MemoryContext oldcontext;
+
/* create a function context for cross-call persistence */
funcctx = SRF_FIRSTCALL_INIT();
+
+ /* Initialize hash table iteration if we have any channels */
+ if (listenChannelsHash != NULL)
+ {
+ oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
+ status = (HASH_SEQ_STATUS *) palloc(sizeof(HASH_SEQ_STATUS));
+ hash_seq_init(status, listenChannelsHash);
+ funcctx->user_fctx = status;
+ MemoryContextSwitchTo(oldcontext);
+ }
+ else
+ {
+ funcctx->user_fctx = NULL;
+ }
}
/* stuff done on every call of the function */
funcctx = SRF_PERCALL_SETUP();
+ status = (HASH_SEQ_STATUS *) funcctx->user_fctx;
- if (funcctx->call_cntr < list_length(listenChannels))
+ if (status != NULL)
{
- char *channel = (char *) list_nth(listenChannels,
- funcctx->call_cntr);
+ struct ChannelHash *entry;
- SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
+ entry = (struct ChannelHash *) hash_seq_search(status);
+ if (entry != NULL)
+ SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(entry->channel));
}
SRF_RETURN_DONE(funcctx);
@@ -894,6 +1077,36 @@ PreCommit_Notify(void)
if (pendingNotifies)
{
ListCell *nextNotify;
+ bool firstIteration = true;
+
+ /*
+ * Build list of unique channels for SignalBackends().
+ *
+ * If we have a channelHashtab, use it to efficiently get the unique
+ * channels. Otherwise, fall back to the linear approach.
+ */
+ pendingNotifyChannels = NIL;
+ if (pendingNotifies->channelHashtab != NULL)
+ {
+ HASH_SEQ_STATUS status;
+ struct ChannelHash *channelEntry;
+
+ hash_seq_init(&status, pendingNotifies->channelHashtab);
+ while ((channelEntry = (struct ChannelHash *) hash_seq_search(&status)) != NULL)
+ pendingNotifyChannels = lappend(pendingNotifyChannels, channelEntry->channel);
+ }
+ else
+ {
+ /* Linear approach for small number of notifications */
+ foreach_ptr(Notification, n, pendingNotifies->events)
+ {
+ char *channel = n->data;
+
+ /* Add if not already in list */
+ if (!list_member_ptr(pendingNotifyChannels, channel))
+ pendingNotifyChannels = lappend(pendingNotifyChannels, channel);
+ }
+ }
/*
* Make sure that we have an XID assigned to the current transaction.
@@ -922,6 +1135,22 @@ PreCommit_Notify(void)
LockSharedObject(DatabaseRelationId, InvalidOid, 0,
AccessExclusiveLock);
+ /*
+ * For the direct advancement optimization in SignalBackends(), we
+ * need to ensure that no other backend can insert queue entries
+ * between queueHeadBeforeWrite and queueHeadAfterWrite. The
+ * heavyweight lock above provides this guarantee, since it serializes
+ * all writers.
+ *
+ * Note: if the heavyweight lock were ever removed for scalability
+ * reasons, we could achieve the same guarantee by holding
+ * NotifyQueueLock in EXCLUSIVE mode across all our insertions, rather
+ * than releasing and reacquiring it for each page as we do below.
+ */
+
+ /* Initialize queueHeadBeforeWrite to a safe default */
+ SET_QUEUE_POS(queueHeadBeforeWrite, 0, 0);
+
/* Now push the notifications into the queue */
nextNotify = list_head(pendingNotifies->events);
while (nextNotify != NULL)
@@ -939,12 +1168,20 @@ PreCommit_Notify(void)
* point in time we can still roll the transaction back.
*/
LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
+ if (firstIteration)
+ {
+ queueHeadBeforeWrite = QUEUE_HEAD;
+ firstIteration = false;
+ }
+
asyncQueueFillWarning();
if (asyncQueueIsFull())
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("too many notifications in the NOTIFY queue")));
nextNotify = asyncQueueAddEntries(nextNotify);
+ queueHeadAfterWrite = QUEUE_HEAD;
+
LWLockRelease(NotifyQueueLock);
}
@@ -957,7 +1194,7 @@ PreCommit_Notify(void)
*
* This is called at transaction commit, after committing to clog.
*
- * Update listenChannels and clear transaction-local state.
+ * Update listenChannelsHash and clear transaction-local state.
*
* If we issued any notifications in the transaction, send signals to
* listening backends (possibly including ourselves) to process them.
@@ -1002,7 +1239,8 @@ AtCommit_Notify(void)
}
/* If no longer listening to anything, get out of listener array */
- if (amRegisteredListener && listenChannels == NIL)
+ if (amRegisteredListener &&
+ (listenChannelsHash == NULL || hash_get_num_entries(listenChannelsHash) == 0))
asyncQueueUnregister();
/*
@@ -1135,50 +1373,145 @@ Exec_ListenPreCommit(void)
static void
Exec_ListenCommit(const char *channel)
{
- MemoryContext oldcontext;
+ ChannelHashKey key;
+ ChannelEntry *entry;
+ bool found;
+ ProcNumber *listeners;
/* Do nothing if we are already listening on this channel */
if (IsListeningOn(channel))
return;
/*
- * Add the new channel name to listenChannels.
+ * Add the new channel name to listenChannelsHash.
*
* XXX It is theoretically possible to get an out-of-memory failure here,
* which would be bad because we already committed. For the moment it
* doesn't seem worth trying to guard against that, but maybe improve this
* later.
*/
- oldcontext = MemoryContextSwitchTo(TopMemoryContext);
- listenChannels = lappend(listenChannels, pstrdup(channel));
- MemoryContextSwitchTo(oldcontext);
+ initListenChannelsHash();
+ (void) hash_search(listenChannelsHash, channel, HASH_ENTER, NULL);
+
+ /* Now update the shared channelHash for SignalBackends() to use */
+ initChannelHash();
+
+ ChannelHashPrepareKey(&key, MyDatabaseId, channel);
+
+ /*
+ * For new entries, we initialize listenersArray to InvalidDsaPointer as a
+ * marker. This handles both the initial creation and potential retry
+ * after OOM.
+ */
+ entry = dshash_find_or_insert(channelHash, &key, &found);
+
+ if (!found)
+ entry->listenersArray = InvalidDsaPointer;
+
+ if (!DsaPointerIsValid(entry->listenersArray))
+ {
+ /* First listener for this channel */
+ entry->listenersArray = dsa_allocate(channelDSA,
+ sizeof(ProcNumber) * INITIAL_LISTENERS_ARRAY_SIZE);
+ entry->numListeners = 0;
+ entry->allocatedListeners = INITIAL_LISTENERS_ARRAY_SIZE;
+ }
+
+ listeners = (ProcNumber *) dsa_get_address(channelDSA,
+ entry->listenersArray);
+
+ for (int i = 0; i < entry->numListeners; i++)
+ {
+ if (listeners[i] == MyProcNumber)
+ {
+ dshash_release_lock(channelHash, entry);
+ return; /* Already registered */
+ }
+ }
+
+ if (entry->numListeners >= entry->allocatedListeners)
+ {
+ int new_size = entry->allocatedListeners * 2;
+ dsa_pointer new_array = dsa_allocate(channelDSA,
+ sizeof(ProcNumber) * new_size);
+ ProcNumber *new_listeners = (ProcNumber *) dsa_get_address(channelDSA,
+ new_array);
+
+ memcpy(new_listeners, listeners,
+ sizeof(ProcNumber) * entry->numListeners);
+
+ dsa_free(channelDSA, entry->listenersArray);
+ entry->listenersArray = new_array;
+ entry->allocatedListeners = new_size;
+ listeners = new_listeners;
+ }
+
+ listeners[entry->numListeners] = MyProcNumber;
+ entry->numListeners++;
+
+ dshash_release_lock(channelHash, entry);
}
/*
* Exec_UnlistenCommit --- subroutine for AtCommit_Notify
*
- * Remove the specified channel name from listenChannels.
+ * Unlisten the specified channel for this backend.
*/
static void
Exec_UnlistenCommit(const char *channel)
{
- ListCell *q;
+ ChannelHashKey key;
+ ChannelEntry *entry;
+ ProcNumber *listeners;
+ int i;
if (Trace_notify)
elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
- foreach(q, listenChannels)
+ /* Remove from our local cache */
+ if (listenChannelsHash != NULL)
+ (void) hash_search(listenChannelsHash, channel, HASH_REMOVE, NULL);
+
+ /* Now remove from the shared channelHash */
+ if (channelHash == NULL)
+ return;
+
+ ChannelHashPrepareKey(&key, MyDatabaseId, channel);
+
+ /* Look up the channel with exclusive lock so we can modify it */
+ entry = dshash_find(channelHash, &key, true);
+ if (entry == NULL)
+ return;
+
+ listeners = (ProcNumber *) dsa_get_address(channelDSA,
+ entry->listenersArray);
+
+ for (i = 0; i < entry->numListeners; i++)
{
- char *lchan = (char *) lfirst(q);
-
- if (strcmp(lchan, channel) == 0)
+ if (listeners[i] == MyProcNumber)
{
- listenChannels = foreach_delete_current(listenChannels, q);
- pfree(lchan);
- break;
+ entry->numListeners--;
+ if (i < entry->numListeners)
+ memmove(&listeners[i], &listeners[i + 1],
+ sizeof(ProcNumber) * (entry->numListeners - i));
+
+ if (entry->numListeners == 0)
+ {
+ /* Last listener for this channel */
+ dsa_free(channelDSA, entry->listenersArray);
+ dshash_delete_entry(channelHash, entry);
+ }
+ else
+ {
+ dshash_release_lock(channelHash, entry);
+ }
+
+ return;
}
}
+ dshash_release_lock(channelHash, entry);
+
/*
* We do not complain about unlistening something not being listened;
* should we?
@@ -1193,34 +1526,68 @@ Exec_UnlistenCommit(const char *channel)
static void
Exec_UnlistenAllCommit(void)
{
+ dshash_seq_status status;
+ ChannelEntry *entry;
+
if (Trace_notify)
elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
- list_free_deep(listenChannels);
- listenChannels = NIL;
+ /* Clear our local cache */
+ if (listenChannelsHash != NULL)
+ {
+ hash_destroy(listenChannelsHash);
+ listenChannelsHash = NULL;
+ }
+
+ /* Now remove from the shared channelHash */
+ if (channelHash == NULL)
+ return;
+
+ dshash_seq_init(&status, channelHash, true);
+ while ((entry = dshash_seq_next(&status)) != NULL)
+ {
+ if (entry->key.dboid == MyDatabaseId)
+ {
+ ProcNumber *listeners;
+ int i;
+
+ listeners = (ProcNumber *) dsa_get_address(channelDSA,
+ entry->listenersArray);
+
+ for (i = 0; i < entry->numListeners; i++)
+ {
+ if (listeners[i] == MyProcNumber)
+ {
+ entry->numListeners--;
+ if (i < entry->numListeners)
+ memmove(&listeners[i], &listeners[i + 1],
+ sizeof(ProcNumber) * (entry->numListeners - i));
+
+ if (entry->numListeners == 0)
+ {
+ dsa_free(channelDSA, entry->listenersArray);
+ dshash_delete_current(&status);
+ }
+ break;
+ }
+ }
+ }
+ }
+ dshash_seq_term(&status);
}
/*
* Test whether we are actively listening on the given channel name.
*
* Note: this function is executed for every notification found in the queue.
- * Perhaps it is worth further optimization, eg convert the list to a sorted
- * array so we can binary-search it. In practice the list is likely to be
- * fairly short, though.
*/
static bool
IsListeningOn(const char *channel)
{
- ListCell *p;
+ if (listenChannelsHash == NULL)
+ return false;
- foreach(p, listenChannels)
- {
- char *lchan = (char *) lfirst(p);
-
- if (strcmp(lchan, channel) == 0)
- return true;
- }
- return false;
+ return (hash_search(listenChannelsHash, channel, HASH_FIND, NULL) != NULL);
}
/*
@@ -1230,7 +1597,7 @@ IsListeningOn(const char *channel)
static void
asyncQueueUnregister(void)
{
- Assert(listenChannels == NIL); /* else caller error */
+ Assert(listenChannelsHash == NULL || hash_get_num_entries(listenChannelsHash) == 0); /* else caller error */
if (!amRegisteredListener) /* nothing to do */
return;
@@ -1242,6 +1609,7 @@ asyncQueueUnregister(void)
/* Mark our entry as invalid */
QUEUE_BACKEND_PID(MyProcNumber) = InvalidPid;
QUEUE_BACKEND_DBOID(MyProcNumber) = InvalidOid;
+ QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
/* and remove it from the list */
if (QUEUE_FIRST_LISTENER == MyProcNumber)
QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyProcNumber);
@@ -1565,12 +1933,15 @@ asyncQueueFillWarning(void)
/*
* Send signals to listening backends.
*
- * Normally we signal only backends in our own database, since only those
- * backends could be interested in notifies we send. However, if there's
- * notify traffic in our database but no traffic in another database that
- * does have listener(s), those listeners will fall further and further
- * behind. Waken them anyway if they're far enough behind, so that they'll
- * advance their queue position pointers, allowing the global tail to advance.
+ * Normally we signal only backends in our own database, that are
+ * listening on the channels with pending notifies, since only those
+ * backends are interested in notifies we send.
+ *
+ * Backends that are known to still be positioned at the queue head
+ * from before our commit can be safely advanced directly to the new
+ * head, since the queue region we wrote is known to contain only our
+ * own notifications. This avoids unnecessary wakeups when there is
+ * nothing of interest to them.
*
* Since we know the ProcNumber and the Pid the signaling is quite cheap.
*
@@ -1583,6 +1954,13 @@ SignalBackends(void)
int32 *pids;
ProcNumber *procnos;
int count;
+ ListCell *lc;
+
+ /*
+ * Attach to the channel hash if needed. We might not have one if this
+ * backend hasn't done LISTEN, but we need it to find listeners.
+ */
+ initChannelHash();
/*
* Identify backends that we need to signal. We don't want to send
@@ -1597,36 +1975,99 @@ SignalBackends(void)
count = 0;
LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
- for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
+ foreach(lc, pendingNotifyChannels)
{
- int32 pid = QUEUE_BACKEND_PID(i);
- QueuePosition pos;
+ char *channel = (char *) lfirst(lc);
+ ChannelEntry *entry = NULL;
+ ProcNumber *listeners;
- Assert(pid != InvalidPid);
- pos = QUEUE_BACKEND_POS(i);
- if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
+ if (channelHash != NULL)
{
- /*
- * Always signal listeners in our own database, unless they're
- * already caught up (unlikely, but possible).
- */
+ ChannelHashKey key;
+
+ ChannelHashPrepareKey(&key, MyDatabaseId, channel);
+ entry = dshash_find(channelHash, &key, false);
+ }
+
+ if (entry == NULL)
+ continue; /* No listeners registered for this channel */
+
+ listeners = (ProcNumber *) dsa_get_address(channelDSA,
+ entry->listenersArray);
+
+ for (int j = 0; j < entry->numListeners; j++)
+ {
+ ProcNumber i = listeners[j];
+ int32 pid;
+ QueuePosition pos;
+
+ if (QUEUE_BACKEND_WAKEUP_PENDING(i))
+ continue;
+
+ pos = QUEUE_BACKEND_POS(i);
+ pid = QUEUE_BACKEND_PID(i);
+
+ /* Skip if caught up */
if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
continue;
+
+ Assert(pid != InvalidPid);
+
+ QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
+ pids[count] = pid;
+ procnos[count] = i;
+ count++;
}
- else
+
+ dshash_release_lock(channelHash, entry);
+ }
+
+ /*
+ * Direct advancement: avoid waking non-caught up backends that aren't
+ * interested in our notifications.
+ */
+ if (pendingNotifies != NULL)
+ {
+ for (ProcNumber i = QUEUE_FIRST_LISTENER;
+ i != INVALID_PROC_NUMBER;
+ i = QUEUE_NEXT_LISTENER(i))
{
+ QueuePosition pos;
+ int32 pid;
+
+ if (QUEUE_BACKEND_WAKEUP_PENDING(i))
+ continue;
+
+ pos = QUEUE_BACKEND_POS(i);
+ pid = QUEUE_BACKEND_PID(i);
+
/*
- * Listeners in other databases should be signaled only if they
- * are far behind.
+ * We need to signal advancing listening backends since we
+ * don't know where they will stop advancing. We also need
+ * to signal listening backends that are idle at a position
+ * before the old queue head since they could be interested
+ * in the messages in-between.
+ *
+ * Listening backends that are not advancing and are stationary at
+ * a position somewhere in the range we just wrote, can safely be
+ * direct advanced to the new queue head, since we know that they
+ * are not interested in our messages.
*/
- if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
- QUEUE_POS_PAGE(pos)) < QUEUE_CLEANUP_DELAY)
- continue;
+ if (QUEUE_BACKEND_IS_ADVANCING(i))
+ {
+ Assert(pid != InvalidPid);
+
+ QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
+ pids[count] = pid;
+ procnos[count] = i;
+ count++;
+ }
+ else if (!QUEUE_POS_PRECEDES(pos, queueHeadBeforeWrite) &&
+ QUEUE_POS_PRECEDES(pos, queueHeadAfterWrite))
+ {
+ QUEUE_BACKEND_POS(i) = queueHeadAfterWrite;
+ }
}
- /* OK, need to signal this one */
- pids[count] = pid;
- procnos[count] = i;
- count++;
}
LWLockRelease(NotifyQueueLock);
@@ -1673,9 +2114,10 @@ AtAbort_Notify(void)
/*
* If we LISTEN but then roll back the transaction after PreCommit_Notify,
* we have registered as a listener but have not made any entry in
- * listenChannels. In that case, deregister again.
+ * listenChannelsHash. In that case, deregister again.
*/
- if (amRegisteredListener && listenChannels == NIL)
+ if (amRegisteredListener &&
+ (listenChannelsHash == NULL || hash_get_num_entries(listenChannelsHash) == 0))
asyncQueueUnregister();
/* And clean up */
@@ -1861,20 +2303,28 @@ asyncQueueReadAllNotifications(void)
AsyncQueueEntry align;
} page_buffer;
- /* Fetch current state */
+ /*
+ * Fetch current state, indicate to others that we have woken up, and that
+ * we now will be advancing our position.
+ */
LWLockAcquire(NotifyQueueLock, LW_SHARED);
/* Assert checks that we have a valid state entry */
Assert(MyProcPid == QUEUE_BACKEND_PID(MyProcNumber));
+ QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
+ head = QUEUE_HEAD;
pos = QUEUE_BACKEND_POS(MyProcNumber);
- head = QUEUE_HEAD;
- LWLockRelease(NotifyQueueLock);
if (QUEUE_POS_EQUAL(pos, head))
{
/* Nothing to do, we have read all notifications already. */
+ LWLockRelease(NotifyQueueLock);
return;
}
+ QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = true;
+
+ LWLockRelease(NotifyQueueLock);
+
/*----------
* Get snapshot we'll use to decide which xacts are still in progress.
* This is trickier than it might seem, because of race conditions.
@@ -1987,6 +2437,8 @@ asyncQueueReadAllNotifications(void)
{
/* Update shared state */
LWLockAcquire(NotifyQueueLock, LW_SHARED);
+
+ QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
QUEUE_BACKEND_POS(MyProcNumber) = pos;
LWLockRelease(NotifyQueueLock);
}
@@ -2186,7 +2638,7 @@ ProcessIncomingNotify(bool flush)
notifyInterruptPending = false;
/* Do nothing else if we aren't actively listening */
- if (listenChannels == NIL)
+ if (listenChannelsHash == NULL || hash_get_num_entries(listenChannelsHash) == 0)
return;
if (Trace_notify)
@@ -2290,13 +2742,15 @@ AddEventToPendingNotifies(Notification *n)
{
Assert(pendingNotifies->events != NIL);
- /* Create the hash table if it's time to */
+ /* Create the hash tables if it's time to */
if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
pendingNotifies->hashtab == NULL)
{
HASHCTL hash_ctl;
ListCell *l;
+ Assert(pendingNotifies->channelHashtab == NULL);
+
/* Create the hash table */
hash_ctl.keysize = sizeof(Notification *);
hash_ctl.entrysize = sizeof(struct NotificationHash);
@@ -2309,10 +2763,22 @@ AddEventToPendingNotifies(Notification *n)
&hash_ctl,
HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
+ /* Create the channel hash table */
+ memset(&hash_ctl, 0, sizeof(hash_ctl));
+ hash_ctl.keysize = NAMEDATALEN;
+ hash_ctl.entrysize = sizeof(struct ChannelHash);
+ hash_ctl.hcxt = CurTransactionContext;
+ pendingNotifies->channelHashtab =
+ hash_create("Pending Notify Channels",
+ 64L,
+ &hash_ctl,
+ HASH_ELEM | HASH_STRINGS | HASH_CONTEXT);
+
/* Insert all the already-existing events */
foreach(l, pendingNotifies->events)
{
Notification *oldn = (Notification *) lfirst(l);
+ char *channel = oldn->data;
bool found;
(void) hash_search(pendingNotifies->hashtab,
@@ -2320,22 +2786,42 @@ AddEventToPendingNotifies(Notification *n)
HASH_ENTER,
&found);
Assert(!found);
+
+ /* Insert channel into channelHashtab */
+ (void) hash_search(pendingNotifies->channelHashtab,
+ channel,
+ HASH_ENTER,
+ &found);
+ /* found may be true if multiple events on same channel */
}
}
/* Add new event to the list, in order */
pendingNotifies->events = lappend(pendingNotifies->events, n);
- /* Add event to the hash table if needed */
+ /* Add event to the hash tables if needed */
if (pendingNotifies->hashtab != NULL)
{
bool found;
+ Assert(pendingNotifies->channelHashtab != NULL);
+
(void) hash_search(pendingNotifies->hashtab,
&n,
HASH_ENTER,
&found);
Assert(!found);
+
+ /* Add channel to channelHashtab */
+ {
+ char *channel = n->data;
+
+ (void) hash_search(pendingNotifies->channelHashtab,
+ channel,
+ HASH_ENTER,
+ &found);
+ /* found may be true if we already have an event on this channel */
+ }
}
}
@@ -2373,7 +2859,7 @@ notification_match(const void *key1, const void *key2, Size keysize)
return 1; /* not equal */
}
-/* Clear the pendingActions and pendingNotifies lists. */
+/* Clear the pendingActions, pendingNotifies, and pendingNotifyChannels lists. */
static void
ClearPendingActionsAndNotifies(void)
{
@@ -2385,6 +2871,7 @@ ClearPendingActionsAndNotifies(void)
*/
pendingActions = NULL;
pendingNotifies = NULL;
+ pendingNotifyChannels = NIL;
}
/*
@@ -2395,3 +2882,16 @@ check_notify_buffers(int *newval, void **extra, GucSource source)
{
return check_slru_buffers("notify_buffers", newval);
}
+
+
+/*
+ * ChannelHashPrepareKey
+ * Prepare a channel key for use as a hash key.
+ */
+static inline void
+ChannelHashPrepareKey(ChannelHashKey *key, Oid dboid, const char *channel)
+{
+ memset(key, 0, sizeof(ChannelHashKey));
+ key->dboid = dboid;
+ strlcpy(key->channel, channel, NAMEDATALEN);
+}
diff --git a/src/backend/utils/activity/wait_event_names.txt b/src/backend/utils/activity/wait_event_names.txt
index c1ac71ff7f2..7c2cf960093 100644
--- a/src/backend/utils/activity/wait_event_names.txt
+++ b/src/backend/utils/activity/wait_event_names.txt
@@ -369,6 +369,7 @@ SubtransBuffer "Waiting for I/O on a sub-transaction SLRU buffer."
MultiXactOffsetBuffer "Waiting for I/O on a multixact offset SLRU buffer."
MultiXactMemberBuffer "Waiting for I/O on a multixact member SLRU buffer."
NotifyBuffer "Waiting for I/O on a <command>NOTIFY</command> message SLRU buffer."
+NotifyChannelHash "Waiting to access the <command>NOTIFY</command> channel hash table."
SerialBuffer "Waiting for I/O on a serializable transaction conflict SLRU buffer."
WALInsert "Waiting to insert WAL data into a memory buffer."
BufferContent "Waiting to access a data page in memory."
diff --git a/src/include/storage/lwlocklist.h b/src/include/storage/lwlocklist.h
index 5b0ce383408..4236965e72a 100644
--- a/src/include/storage/lwlocklist.h
+++ b/src/include/storage/lwlocklist.h
@@ -101,6 +101,7 @@ PG_LWLOCKTRANCHE(SUBTRANS_BUFFER, SubtransBuffer)
PG_LWLOCKTRANCHE(MULTIXACTOFFSET_BUFFER, MultiXactOffsetBuffer)
PG_LWLOCKTRANCHE(MULTIXACTMEMBER_BUFFER, MultiXactMemberBuffer)
PG_LWLOCKTRANCHE(NOTIFY_BUFFER, NotifyBuffer)
+PG_LWLOCKTRANCHE(NOTIFY_CHANNEL_HASH, NotifyChannelHash)
PG_LWLOCKTRANCHE(SERIAL_BUFFER, SerialBuffer)
PG_LWLOCKTRANCHE(WAL_INSERT, WALInsert)
PG_LWLOCKTRANCHE(BUFFER_CONTENT, BufferContent)
diff --git a/src/tools/pgindent/typedefs.list b/src/tools/pgindent/typedefs.list
index 432509277c9..ee3540643c8 100644
--- a/src/tools/pgindent/typedefs.list
+++ b/src/tools/pgindent/typedefs.list
@@ -412,6 +412,8 @@ CatalogIdMapEntry
CatalogIndexState
ChangeVarNodes_callback
ChangeVarNodes_context
+ChannelEntry
+ChannelHashKey
CheckPoint
CheckPointStmt
CheckpointStatsData
--
2.50.1
reply
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: [email protected]
Cc: [email protected]
Subject: Re: Optimize LISTEN/NOTIFY
In-Reply-To: <[email protected]>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox