agora inbox for [email protected]  

help / color / mirror / Atom feed

pglife and post-beta commits
522+ messages / 3 participants
[nested] [flat]

* pglife and post-beta commits
@ 2018-06-19 22:18 Bruce Momjian <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Bruce Momjian @ 2018-06-19 22:18 UTC (permalink / raw)
  To: pgsql-hackers

FYI, you can now see all the commits since the most recent beta in
pglife (http://pglife.momjian.us/).  It is the plus sign link after "11
beta".  The link currently is:

	https://git.postgresql.org/gitweb/?p=postgresql.git;a=shortlog;h=refs/heads/master;hp=586e4e6df5b85d...

-- 
  Bruce Momjian  <[email protected]>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

+ As you are, so once was I.  As I am, so you will be. +
+                      Ancient Roman grave inscription +




^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v37 11/11] Add documentations about Incremental View Maintenance
@ 2019-12-20 01:25 Yugo Nagata <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Yugo Nagata @ 2019-12-20 01:25 UTC (permalink / raw)

---
 doc/src/sgml/catalogs.sgml                    |   9 +
 .../sgml/ref/create_materialized_view.sgml    | 124 ++++-
 .../sgml/ref/refresh_materialized_view.sgml   |   8 +-
 doc/src/sgml/rules.sgml                       | 437 ++++++++++++++++++
 doc/src/sgml/system-views.sgml                |   9 +
 5 files changed, 583 insertions(+), 4 deletions(-)

diff --git a/doc/src/sgml/catalogs.sgml b/doc/src/sgml/catalogs.sgml
index 4b474c13917..9b1c88161a7 100644
--- a/doc/src/sgml/catalogs.sgml
+++ b/doc/src/sgml/catalogs.sgml
@@ -2276,6 +2276,15 @@ SCRAM-SHA-256$<replaceable>&lt;iteration count&gt;</replaceable>:<replaceable>&l
       </para></entry>
      </row>
 
+     <row>
+      <entry role="catalog_table_entry"><para role="column_definition">
+       <structfield>relisivm</structfield> <type>bool</type>
+      </para>
+      <para>
+       True if relation is incrementally maintainable materialized view
+      </para></entry>
+     </row>
+
      <row>
       <entry role="catalog_table_entry"><para role="column_definition">
        <structfield>relrewrite</structfield> <type>oid</type>
diff --git a/doc/src/sgml/ref/create_materialized_view.sgml b/doc/src/sgml/ref/create_materialized_view.sgml
index 62d897931c3..f49db209558 100644
--- a/doc/src/sgml/ref/create_materialized_view.sgml
+++ b/doc/src/sgml/ref/create_materialized_view.sgml
@@ -21,7 +21,7 @@ PostgreSQL documentation
 
  <refsynopsisdiv>
 <synopsis>
-CREATE MATERIALIZED VIEW [ IF NOT EXISTS ] <replaceable>table_name</replaceable>
+CREATE [ INCREMENTAL ] MATERIALIZED VIEW [ IF NOT EXISTS ] <replaceable>table_name</replaceable>
     [ (<replaceable>column_name</replaceable> [, ...] ) ]
     [ USING <replaceable class="parameter">method</replaceable> ]
     [ WITH ( <replaceable class="parameter">storage_parameter</replaceable> [= <replaceable class="parameter">value</replaceable>] [, ... ] ) ]
@@ -60,6 +60,125 @@ CREATE MATERIALIZED VIEW [ IF NOT EXISTS ] <replaceable>table_name</replaceable>
   <title>Parameters</title>
 
   <variablelist>
+   <varlistentry>
+    <term><literal>INCREMENTAL</literal></term>
+    <listitem>
+     <para>
+      If specified, some triggers are automatically created so that the rows
+      of the materialized view are immediately updated when base tables of the
+      materialized view are updated. In general, this allows faster update of
+      the materialized view at a price of slower update of the base tables
+      because the triggers will be invoked. We call this form of materialized
+      view as "Incrementally Maintainable Materialized View" (IMMV).
+     </para>
+     <para>
+      When <acronym>IMMV</acronym> is defined without using <command>WITH NO DATA</command>,
+      a unique index is created on the view automatically if possible.  If the view
+      definition query has a GROUP BY clause, a unique index is created on the columns
+      of GROUP BY expressions.  Also, if the view has DISTINCT clause, a unique index
+      is created on all columns in the target list.  Otherwise, if the view contains all
+      primary key attritubes of its base tables in the target list, a unique index is
+      created on these attritubes.  In other cases, no index is created.
+     </para>
+     <para>
+      There are restrictions of query definitions allowed to use this
+      option. The following are supported in query definitions for IMMV:
+      <itemizedlist>
+
+       <listitem>
+        <para>
+         Inner joins (including self-joins).
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         Some built-in aggregate functions (count, sum, avg, min, max) without a HAVING
+         clause.
+        </para>
+        </listitem>
+      </itemizedlist>
+
+      Unsupported queries with this option include the following:
+
+      <itemizedlist>
+       <listitem>
+        <para>
+         Outer joins.
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         Sub-queries.
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         Aggregate functions other than built-in count, sum, avg, min and max.
+        </para>
+       </listitem>
+       <listitem>
+        <para>
+         Aggregate functions with a HAVING clause.
+        </para>
+       </listitem>
+       <listitem>
+        <para>
+         DISTINCT ON, WINDOW, VALUES, LIMIT and OFFSET clause.
+        </para>
+       </listitem>
+      </itemizedlist>
+
+      Other restrictions include:
+      <itemizedlist>
+
+       <listitem>
+        <para>
+         IMMVs must be based on simple base tables. It's not supported to
+         create them on top of views or materialized views.
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         It is not supported to include system columns in an IMMV.
+         <programlisting>
+CREATE INCREMENTAL MATERIALIZED VIEW  mv_ivm02 AS SELECT i,j FROM mv_base_a WHERE xmin = '610';
+ERROR:  system column is not supported with IVM
+         </programlisting>
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         Non-immutable functions are not supported.
+         <programlisting>
+CREATE INCREMENTAL MATERIALIZED VIEW  mv_ivm12 AS SELECT i,j FROM mv_base_a WHERE i = random()::int;
+ERROR:  functions in IMMV must be marked IMMUTABLE
+         </programlisting>
+        </para>
+        </listitem>
+
+       <listitem>
+        <para>
+         IMMVs do not support expressions that contains aggregates
+        </para>
+       </listitem>
+
+       <listitem>
+        <para>
+         Logical replication does not support IMMVs.
+        </para>
+       </listitem>
+
+      </itemizedlist>
+
+     </para>
+    </listitem>
+   </varlistentry>
+
    <varlistentry>
     <term><literal>IF NOT EXISTS</literal></term>
     <listitem>
@@ -157,7 +276,8 @@ CREATE MATERIALIZED VIEW [ IF NOT EXISTS ] <replaceable>table_name</replaceable>
       This clause specifies whether or not the materialized view should be
       populated at creation time.  If not, the materialized view will be
       flagged as unscannable and cannot be queried until <command>REFRESH
-      MATERIALIZED VIEW</command> is used.
+      MATERIALIZED VIEW</command> is used.  Also, if the view is IMMV,
+      triggers for maintaining the view are not created.
      </para>
     </listitem>
    </varlistentry>
diff --git a/doc/src/sgml/ref/refresh_materialized_view.sgml b/doc/src/sgml/ref/refresh_materialized_view.sgml
index 8ed43ade803..a4d729bdf0f 100644
--- a/doc/src/sgml/ref/refresh_materialized_view.sgml
+++ b/doc/src/sgml/ref/refresh_materialized_view.sgml
@@ -36,9 +36,13 @@ REFRESH MATERIALIZED VIEW [ CONCURRENTLY ] <replaceable class="parameter">name</
    privilege on the materialized view.  The old contents are discarded.  If
    <literal>WITH DATA</literal> is specified (or defaults) the backing query
    is executed to provide the new data, and the materialized view is left in a
-   scannable state.  If <literal>WITH NO DATA</literal> is specified no new
+   scannable state.  If the view is an incrementally maintainable materialized
+   view (IMMV) and was unpopulated, triggers for maintaining the view are
+   created. Also, a unique index is created for IMMV if it is possible and the
+   view doesn't have that yet.
+   If <literal>WITH NO DATA</literal> is specified no new
    data is generated and the materialized view is left in an unscannable
-   state.
+   state.  If the view is IMMV, the triggers are dropped.
   </para>
   <para>
    <literal>CONCURRENTLY</literal> and <literal>WITH NO DATA</literal> may not
diff --git a/doc/src/sgml/rules.sgml b/doc/src/sgml/rules.sgml
index 7f23962f524..da9ad3c6316 100644
--- a/doc/src/sgml/rules.sgml
+++ b/doc/src/sgml/rules.sgml
@@ -1103,6 +1103,443 @@ SELECT word FROM words ORDER BY word &lt;-&gt; 'caterpiler' LIMIT 10;
 
 </sect1>
 
+<sect1 id="rules-ivm">
+<title>Incremental View Maintenance</title>
+
+<indexterm zone="rules-ivm">
+ <primary>incremental view maintenance</primary>
+</indexterm>
+
+<indexterm zone="rules-ivm">
+ <primary>materialized view</primary>
+ <secondary>incremental view maintenance</secondary>
+</indexterm>
+
+<indexterm zone="rules-ivm">
+ <primary>view</primary>
+ <secondary>incremental view maintenance</secondary>
+</indexterm>
+
+<sect2 id="rules-ivm-overview">
+<title>Overview</title>
+
+<para>
+    Incremental View Maintenance (<acronym>IVM</acronym>) is a way to make
+    materialized views up-to-date in which only incremental changes are computed
+    and applied on views rather than recomputing the contents from scratch as
+    <command>REFRESH MATERIALIZED VIEW</command> does.  <acronym>IVM</acronym>
+    can update materialized views more efficiently than recomputation when only
+    small parts of the view are changed.
+</para>
+
+<para>
+    There are two approaches with regard to timing of view maintenance:
+    immediate and deferred.  In immediate maintenance, views are updated in the
+    same transaction that its base table is modified.  In deferred maintenance,
+    views are updated after the transaction is committed, for example, when the
+    view is accessed, as a response to user command like <command>REFRESH
+    MATERIALIZED VIEW</command>, or periodically in background, and so on.
+    <productname>PostgreSQL</productname> currently implements only a kind of
+    immediate maintenance, in which materialized views are updated immediately
+    in AFTER triggers when a base table is modified.
+</para>
+
+<para>
+    To create materialized views supporting <acronym>IVM</acronym>, use the
+    <command>CREATE INCREMENTAL MATERIALIZED VIEW</command>, for example:
+<programlisting>
+CREATE <emphasis>INCREMENTAL</emphasis> MATERIALIZED VIEW mymatview AS SELECT * FROM mytab;
+</programlisting>
+    When a materialized view is created with the <literal>INCREMENTAL</literal>
+    keyword, some triggers are automatically created so that the view's contents are
+    immediately updated when its base tables are modified. We call this form
+    of materialized view an Incrementally Maintainable Materialized View
+    (<acronym>IMMV</acronym>).
+<programlisting>
+postgres=# CREATE INCREMENTAL MATERIALIZED VIEW m AS SELECT * FROM t0;
+NOTICE:  could not create an index on materialized view "m" automatically
+HINT:  Create an index on the materialized view for effcient incremental maintenance.
+SELECT 3
+postgres=# SELECT * FROM m;
+ i
+---
+ 1
+ 2
+ 3
+(3 rows)
+
+postgres=# INSERT INTO t0 VALUES (4);
+INSERT 0 1
+postgres=# SELECT * FROM m; -- automatically updated
+ i
+---
+ 1
+ 2
+ 3
+ 4
+(4 rows)
+</programlisting>
+</para>
+
+<para>
+    Some <acronym>IMMV</acronym>s have hidden columns which are added
+    automatically when a materialized view is created. Their name starts
+    with <literal>__ivm_</literal> and they contain information required
+    for maintaining the <acronym>IMMV</acronym>. Such columns are not visible
+    when the <acronym>IMMV</acronym> is accessed by <literal>SELECT *</literal>
+    but are visible if the column name is explicitly specified in the target
+    list. We can also see the hidden columns in <literal>\d</literal>
+    meta-commands of <command>psql</command> commands.
+</para>
+
+<para>
+    In general, <acronym>IMMV</acronym>s allow faster updates of materialized
+    views at the price of slower updates to their base tables. Updates of
+    <acronym>IMMV</acronym> is slower because triggers will be invoked and the
+    view is updated in triggers per modification statement.
+</para>
+
+<para>
+    For example, suppose a normal materialized view defined as below:
+
+<programlisting>
+test=# CREATE INCREMENTAL MATERIALIZED VIEW mv_ivm AS
+        SELECT a.aid, b.bid, a.abalance, b.bbalance
+        FROM pgbench_accounts a JOIN pgbench_branches b USING(bid);
+SELECT 10000000
+
+</programlisting>
+
+    Updating a tuple in a base table of this materialized view is rapid but the
+   <command>REFRESH MATERIALIZED VIEW</command> command on this view takes a long time:
+
+<programlisting>
+test=# UPDATE pgbench_accounts SET abalance = 1000 WHERE aid = 1;
+UPDATE 1
+Time: 0.990 ms
+
+test=# REFRESH MATERIALIZED VIEW mv_normal ;
+REFRESH MATERIALIZED VIEW
+Time: 33533.952 ms (00:33.534)
+</programlisting>
+</para>
+
+<para>
+    On the other hand, after creating <acronym>IMMV</acronym> with the same view
+    definition as below:
+
+<programlisting>
+test=# CREATE INCREMENTAL MATERIALIZED VIEW mv_ivm AS
+        SELECT a.aid, b.bid, a.abalance, b.bbalance
+        FROM pgbench_accounts a JOIN pgbench_branches b USING(bid);
+test=# UPDATE pgbench_accounts SET abalance = 1000 WHERE aid = 1;
+NOTICE:  created index "mv_ivm_index" on materialized view "mv_ivm"
+</programlisting>
+
+    updating a tuple in a base table takes more than the normal view,
+    but its content is updated automatically and this is faster than the
+    <command>REFRESH MATERIALIZED VIEW</command> command.
+
+<programlisting>
+test=# UPDATE pgbench_accounts SET abalance = 1000 WHERE aid = 1;
+UPDATE 1
+Time: 13.068 ms
+</programlisting>
+
+</para>
+
+<para>
+    Appropriate indexes on <acronym>IMMV</acronym>s are necessary for
+    efficient <acronym>IVM</acronym> because it looks for tuples to be
+    updated in <acronym>IMMV</acronym>.  If there are no indexes, it
+    will take a long time.
+</para>
+
+<para>
+    Therefore, when <acronym>IMMV</acronym> is defined, a unique index is created on the view
+    automatically if possible.  If the view definition query has a GROUP BY clause, a unique
+    index is created on the columns of GROUP BY expressions.  Also, if the view has DISTINCT
+    clause, a unique index is created on all columns in the target list. Otherwise, if the
+    view contains all primary key attritubes of its base tables in the target list, a unique
+    index is created on these attritubes.  In other cases, no index is created.
+</para>
+
+<para>
+    In the previous example, a unique index "mv_ivm_index" is created on aid and bid
+    columns of materialized view "mv_ivm", and this enables the rapid update of the view.
+    Dropping this index make updating the view take a loger time.
+<programlisting>
+test=# DROP INDEX mv_ivm_index;
+DROP INDEX
+Time: 67.081 ms
+
+test=# UPDATE pgbench_accounts SET abalance = 1000 WHERE aid = 1;
+UPDATE 1
+Time: 16386.245 ms (00:16.386)
+</programlisting>
+
+</para>
+
+<para>
+    <acronym>IVM</acronym> is effective when we want to keep a materialized
+    view up-to-date and small fraction of a base table is modified
+    infrequently.  Due to the overhead of immediate maintenance, <acronym>IVM</acronym>
+    is not effective when a base table is modified frequently.  Also, when a
+    large part of a base table is modified or large data is inserted into a
+    base table, <acronym>IVM</acronym> is not effective and the cost of
+    maintenance can be larger than the <command>REFRESH MATERIALIZED VIEW</command>
+    command. In such situation, we can use <command>REFRESH MATERIALIZED VIEW</command>
+    and specify <literal>WITH NO DATA</literal> to disable immediate
+    maintenance before modifying a base table. After a base table modification,
+    execute the <command>REFRESH MATERIALIZED VIEW</command> (with <literal>WITH DATA</literal>)
+    command to refresh the view data and enable immediate maintenance.
+</para>
+
+</sect2>
+
+<sect2 id="rules-ivm-support">
+<title>Supported View Definitions and Restrictions</title>
+
+<para>
+    Currently, we can create <acronym>IMMV</acronym>s using inner joins, and some
+    aggregates. However, several restrictions apply to the definition of IMMV.
+</para>
+
+<sect3 id="rules-ivm-support-joins">
+<title>Joins</title>
+<para>
+    Inner joins including self-join are supported. Outer joins are not supported.
+</para>
+</sect3>
+
+<sect3 id="rules-ivm-support-aggregates">
+<title>Aggregates</title>
+<para>
+    Supported aggregate functions are <function>count</function>, <function>sum</function>,
+    <function>avg</function>, <function>min</function>, and <function>max</function>.
+    Currently, only built-in aggregate functions are supported and user defined
+    aggregates cannot be used.  When a base table is modified, the new aggregated
+    values are incrementally calculated using the old aggregated values and values
+    of related hidden columns stored in <acronym>IMMV</acronym>.
+</para>
+
+<para>
+     Note that for <function>min</function> or <function>max</function>, the new values
+     could be re-calculated from base tables with regard to the affected groups when a
+     tuple containing the current minimal or maximal values are deleted from a base table.
+     Therefore, it can takes a long time to update an <acronym>IMMV</acronym> containing
+     these functions.
+</para>
+
+<para>
+    Also note that using <function>sum</function> or <function>avg</function> on
+    <type>real</type> (<type>float4</type>) type or <type>double precision</type>
+    (<type>float8</type>) type in <acronym>IMMV</acronym> is unsafe. This is
+    because aggregated values in <acronym>IMMV</acronym> can become different from
+    results calculated from base tables due to the limited precision of these types.
+    To avoid this problem, use the <type>numeric</type> type instead.
+</para>
+
+    <sect4 id="rules-ivm-restrictions-aggregates">
+    <title>Restrictions on Aggregates</title>
+    <para>
+        There are the following restrictions:
+    <itemizedlist>
+        <listitem>
+        <para>
+            If we have a <literal>GROUP BY</literal> clause, expressions specified in
+           <literal>GROUP BY</literal> must appear in the target list.  This is
+           how tuples to be updated in the <acronym>IMMV</acronym> are identified.
+           These attributes are used as scan keys for searching tuples in the
+           <acronym>IMMV</acronym>, so indexes on them are required for efficient
+           <acronym>IVM</acronym>.
+        </para>
+        </listitem>
+
+        <listitem>
+        <para>
+            <literal>HAVING</literal> clause cannot be used.
+        </para>
+        </listitem>
+    </itemizedlist>
+    </para>
+    </sect4>
+</sect3>
+
+<sect3 id="rules-ivm-general-restricitons">
+<title>Other General Restrictions</title>
+<para>
+    There are other restrictions which generally apply to <acronym>IMMV</acronym>:
+    <itemizedlist>
+        <listitem>
+          <para>
+           Sub-queries cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+           CTEs cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+           Window functions cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            <acronym>IMMV</acronym>s must be based on simple base tables.  It's not
+               supported to create them on top of views, materialized views, foreign tables, inhe.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            LIMIT and OFFSET clauses cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            <acronym>IMMV</acronym>s cannot contain system columns.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            <acronym>IMMV</acronym>s cannot contain non-immutable functions.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            UNION/INTERSECT/EXCEPT clauses cannnot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            DISTINCT ON clauses cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            TABLESAMPLE parameter cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            inheritance parent tables cannnot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            VALUES clause cannnot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            <literal>GROUPING SETS</literal> and <literal>FILTER</literal> clauses cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            FOR UPDATE/SHARE cannot be used.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            targetlist cannot contain columns whose name start with <literal>__ivm_</literal>.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+            targetlist cannot contain expressions which contain an aggregate in it.
+          </para>
+        </listitem>
+
+        <listitem>
+          <para>
+              Logical replication is not supported, that is, even when a base table
+               at a publisher node is modified, <acronym>IMMV</acronym>s at subscriber
+               nodes are not updated.
+          </para>
+        </listitem>
+
+    </itemizedlist>
+</para>
+</sect3>
+
+</sect2>
+
+<sect2 id="rules-ivm-distinct">
+<title><literal>DISTINCT</literal></title>
+
+<para>
+    <productname>PostgreSQL</productname> supports <acronym>IMMV</acronym> with
+    <literal>DISTINCT</literal>.  For example, suppose a <acronym>IMMV</acronym>
+    defined with <literal>DISTINCT</literal> on a base table containing duplicate
+    tuples.  When tuples are deleted from the base table, a tuple in the view is
+    deleted if and only if the multiplicity of the tuple becomes zero.  Moreover,
+    when tuples are inserted into the base table, a tuple is inserted into the
+    view only if the same tuple doesn't already exist in it.
+</para>
+
+<para>
+    Physically, an <acronym>IMMV</acronym> defined with <literal>DISTINCT</literal>
+    contains tuples after eliminating duplicates, and the multiplicity of each tuple
+    is stored in a hidden column named <literal>__ivm_count__</literal>.
+</para>
+</sect2>
+
+<sect2 id="rules-ivm-concurrent-transactions">
+<title>Concurrent Transactions</title>
+<para>
+    Suppose an <acronym>IMMV</acronym> is defined on two base tables and each
+    table was modified in different a concurrent transaction simultaneously.
+    In the transaction which was committed first, <acronym>IMMV</acronym> can
+    be updated considering only the change which happened in this transaction.
+    On the other hand, in order to update the view correctly in the transaction
+    which was committed later, we need to know the changes occurred in
+    both transactions.  For this reason, <literal>ExclusiveLock</literal>
+    is held on an <acronym>IMMV</acronym> immediately after a base table is
+    modified in <literal>READ COMMITTED</literal> mode to make sure that
+    the <acronym>IMMV</acronym> is updated in the latter transaction after
+    the former transaction is committed.  In <literal>REPEATABLE READ</literal>
+    or <literal>SERIALIZABLE</literal> mode, an error is raised immediately
+    if lock acquisition fails because any changes which occurred in
+    other transactions are not be visible in these modes and
+    <acronym>IMMV</acronym> cannot be updated correctly in such situations.
+    However, as an exception if the view has only one base table and
+    <command>INSERT</command> is performed on the table,
+    the lock held on thew view is <literal>RowExclusiveLock</literal>.
+</para>
+</sect2>
+
+<sect2 id="rules-ivm-rls">
+<title>Row Level Security</title>
+<para>
+    If some base tables have row level security policy, rows that are not visible
+    to the materialized view's owner are excluded from the result.  In addition, such
+    rows are excluded as well when views are incrementally maintained.  However, if a
+    new policy is defined or policies are changed after the materialized view was created,
+    the new policy will not be applied to the view contents.  To apply the new policy,
+    you need to refresh materialized views.
+</para>
+</sect2>
+
+</sect1>
+
 <sect1 id="rules-update">
 <title>Rules on <command>INSERT</command>, <command>UPDATE</command>, and <command>DELETE</command></title>
 
diff --git a/doc/src/sgml/system-views.sgml b/doc/src/sgml/system-views.sgml
index 2ebec6928d5..65d1e2041d2 100644
--- a/doc/src/sgml/system-views.sgml
+++ b/doc/src/sgml/system-views.sgml
@@ -2231,6 +2231,15 @@ SELECT * FROM pg_locks pl LEFT JOIN pg_prepared_xacts ppx
       </para></entry>
      </row>
 
+     <row>
+      <entry role="catalog_table_entry"><para role="column_definition">
+       <structfield>isimmv</structfield> <type>bool</type>
+      </para>
+      <para>
+       True if materialized view is incrementally maintainable
+      </para></entry>
+     </row>
+
      <row>
       <entry role="catalog_table_entry"><para role="column_definition">
        <structfield>definition</structfield> <type>text</type>
-- 
2.43.0


--Multipart=_Fri__29_May_2026_23_14_17_+0900_Te0o73X2VqYK57Gd
Content-Type: text/x-diff;
 name="v37-0010-Add-regression-tests-for-Incremental-View-Mainte.patch"
Content-Disposition: attachment;
 filename="v37-0010-Add-regression-tests-for-Incremental-View-Mainte.patch"
Content-Transfer-Encoding: 7bit



^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread

* [PATCH v1] bufmgr: aio: Prototype for not waiting for already-in-progress IO
@ 2025-08-15 15:01 Andres Freund <[email protected]>
  0 siblings, 0 replies; 522+ messages in thread

From: Andres Freund @ 2025-08-15 15:01 UTC (permalink / raw)

Author:
Reviewed-by:
Discussion: https://postgr.es/m/
Backpatch:
---
 src/include/storage/bufmgr.h        |   1 +
 src/backend/storage/buffer/bufmgr.c | 133 ++++++++++++++++++++++++++--
 2 files changed, 125 insertions(+), 9 deletions(-)

diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index 41fdc1e7693..7ddb867bc99 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -137,6 +137,7 @@ struct ReadBuffersOperation
 	int			flags;
 	int16		nblocks;
 	int16		nblocks_done;
+	bool		foreign_io;
 	PgAioWaitRef io_wref;
 	PgAioReturn io_return;
 };
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index fd7e21d96d3..de755fd53ad 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1557,6 +1557,41 @@ ReadBuffersCanStartIOOnce(Buffer buffer, bool nowait)
 		return StartBufferIO(GetBufferDescriptor(buffer - 1), true, nowait);
 }
 
+static inline bool
+ReadBuffersIOAlreadyInProgress(ReadBuffersOperation *operation, Buffer buffer)
+{
+	BufferDesc *desc;
+	uint32		buf_state;
+	PgAioWaitRef iow;
+
+	pgaio_wref_clear(&iow);
+
+	if (BufferIsLocal(buffer))
+	{
+		desc = GetLocalBufferDescriptor(-buffer - 1);
+		buf_state = pg_atomic_read_u32(&desc->state);
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+	}
+	else
+	{
+		desc = GetBufferDescriptor(buffer - 1);
+		buf_state = LockBufHdr(desc);
+
+		if (buf_state & BM_IO_IN_PROGRESS)
+			iow = desc->io_wref;
+		UnlockBufHdr(desc, buf_state);
+	}
+
+	if (pgaio_wref_valid(&iow))
+	{
+		operation->io_wref = iow;
+		return true;
+	}
+
+	return false;
+}
+
 /*
  * Helper for AsyncReadBuffers that tries to get the buffer ready for IO.
  */
@@ -1689,7 +1724,7 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 			 *
 			 * we first check if we already know the IO is complete.
 			 */
-			if (aio_ret->result.status == PGAIO_RS_UNKNOWN &&
+			if ((operation->foreign_io || aio_ret->result.status == PGAIO_RS_UNKNOWN) &&
 				!pgaio_wref_check_done(&operation->io_wref))
 			{
 				instr_time	io_start = pgstat_prepare_io_time(track_io_timing);
@@ -1708,11 +1743,38 @@ WaitReadBuffers(ReadBuffersOperation *operation)
 				Assert(pgaio_wref_check_done(&operation->io_wref));
 			}
 
-			/*
-			 * We now are sure the IO completed. Check the results. This
-			 * includes reporting on errors if there were any.
-			 */
-			ProcessReadBuffersResult(operation);
+			if (operation->foreign_io)
+			{
+				Buffer		buffer = operation->buffers[operation->nblocks_done];
+				BufferDesc *desc;
+				uint32		buf_state;
+
+				if (BufferIsLocal(buffer))
+				{
+					desc = GetLocalBufferDescriptor(-buffer - 1);
+					buf_state = pg_atomic_read_u32(&desc->state);
+				}
+				else
+				{
+					desc = GetBufferDescriptor(buffer - 1);
+					buf_state = LockBufHdr(desc);
+					UnlockBufHdr(desc, buf_state);
+				}
+
+				if (buf_state & BM_VALID)
+				{
+					operation->nblocks_done += 1;
+					Assert(operation->nblocks_done <= operation->nblocks);
+				}
+			}
+			else
+			{
+				/*
+				 * We now are sure the IO completed. Check the results. This
+				 * includes reporting on errors if there were any.
+				 */
+				ProcessReadBuffersResult(operation);
+			}
 		}
 
 		/*
@@ -1798,6 +1860,56 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		io_object = IOOBJECT_RELATION;
 	}
 
+	/*
+	 * If AIO is in progress, be it in this backend or another backend, we
+	 * just associate the wait reference with the operation and wait in
+	 * WaitReadBuffers(). This turns out to be important for performance in
+	 * two workloads:
+	 *
+	 * 1) A read stream that has to read the same block multiple times within
+	 *	  the readahead distance. This can happen e.g. for the table accesses
+	 *	  of an index scan.
+	 *
+	 * 2) Concurrent scans by multiple backends on the same relation.
+	 *
+	 * If we were to synchronously wait for the in-progress IO, we'd not be
+	 * able to keep enough I/O in flight.
+	 *
+	 *
+	 * If we do find there is ongoing I/O for the buffer, we set up a 1-block
+	 * ReadBuffersOperation that WaitReadBuffers then can wait on.
+	 */
+	if (1 && ReadBuffersIOAlreadyInProgress(operation, buffers[nblocks_done]))
+	{
+		/* FIXME: probably need to wait if io_method == sync? */
+
+		*nblocks_progress = 1;
+		did_start_io = true;
+		operation->foreign_io = true;
+
+		if (0)
+			elog(LOG, "using foreign IO path");
+
+		/* FIXME: trace point */
+
+		/*
+		 * FIXME: how should this be accounted for in stats? Account as a hit
+		 * for now, quite likely *we* started this IO.
+		 */
+		if (persistence == RELPERSISTENCE_TEMP)
+			pgBufferUsage.local_blks_hit += 1;
+		else
+			pgBufferUsage.shared_blks_hit += 1;
+
+		if (operation->rel)
+			pgstat_count_buffer_hit(operation->rel);
+		pgstat_count_io_op(io_object, io_context, IOOP_HIT, 1, 0);
+		if (VacuumCostActive)
+			VacuumCostBalance += VacuumCostPageHit;
+
+		return true;
+	}
+
 	/*
 	 * If zero_damaged_pages is enabled, add the READ_BUFFERS_ZERO_ON_ERROR
 	 * flag. The reason for that is that, hopefully, zero_damaged_pages isn't
@@ -1855,9 +1967,9 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 	/*
 	 * Check if we can start IO on the first to-be-read buffer.
 	 *
-	 * If an I/O is already in progress in another backend, we want to wait
-	 * for the outcome: either done, or something went wrong and we will
-	 * retry.
+	 * If a synchronous I/O in progress in another backend (it can't be this
+	 * backend), we want to wait for the outcome: either done, or something
+	 * went wrong and we will retry.
 	 */
 	if (!ReadBuffersCanStartIO(buffers[nblocks_done], false))
 	{
@@ -1970,6 +2082,7 @@ AsyncReadBuffers(ReadBuffersOperation *operation, int *nblocks_progress)
 		if (VacuumCostActive)
 			VacuumCostBalance += VacuumCostPageMiss * io_buffers_len;
 
+		operation->foreign_io = false;
 		*nblocks_progress = io_buffers_len;
 		did_start_io = true;
 	}
@@ -5986,6 +6099,8 @@ WaitIO(BufferDesc *buf)
 		 */
 		if (pgaio_wref_valid(&iow))
 		{
+			if (0)
+				elog(LOG, "foreign wait");
 			pgaio_wref_wait(&iow);
 
 			/*
-- 
2.48.1.76.g4e746b1a31.dirty


--j4imdtjxqt5tb6mo--





^ permalink  raw  reply  [nested|flat] 522+ messages in thread