From 4087138638a20b07215a87575d5d6a6699a4a4da Mon Sep 17 00:00:00 2001 From: Peter Geoghegan Date: Wed, 19 Aug 2026 18:01:40 -0400 Subject: [PATCH 1/2] Add repro: GIN VACUUM unlinks a live page past an unfinished split. An unfinished page split leaves the new right half reachable through the sibling chain, but with no downlink in its parent. VACUUM deletes empty pages by walking downlinks, so it mistakes the incompletely split page for the left sibling of the page it is deleting, and relinks the chain over the top of the live right half. Index scans then silently miss every row indexed there. gin_index_check() fails to detect the corruption. Co-Authored-By: Claude Opus 5 (1M context) --- src/test/modules/gin/Makefile | 4 +- .../gin/expected/gin_vacuum_orphan_page.out | 119 ++++++++++++++++++ src/test/modules/gin/meson.build | 1 + .../gin/sql/gin_vacuum_orphan_page.sql | 96 ++++++++++++++ 4 files changed, 218 insertions(+), 2 deletions(-) create mode 100644 src/test/modules/gin/expected/gin_vacuum_orphan_page.out create mode 100644 src/test/modules/gin/sql/gin_vacuum_orphan_page.sql diff --git a/src/test/modules/gin/Makefile b/src/test/modules/gin/Makefile index e007e38ac..33159bc07 100644 --- a/src/test/modules/gin/Makefile +++ b/src/test/modules/gin/Makefile @@ -1,8 +1,8 @@ # src/test/modules/gin/Makefile -EXTRA_INSTALL = src/test/modules/injection_points +EXTRA_INSTALL = src/test/modules/injection_points contrib/pageinspect -REGRESS = gin_incomplete_splits +REGRESS = gin_incomplete_splits gin_vacuum_orphan_page ifdef USE_PGXS PG_CONFIG = pg_config diff --git a/src/test/modules/gin/expected/gin_vacuum_orphan_page.out b/src/test/modules/gin/expected/gin_vacuum_orphan_page.out new file mode 100644 index 000000000..b47a75102 --- /dev/null +++ b/src/test/modules/gin/expected/gin_vacuum_orphan_page.out @@ -0,0 +1,119 @@ +-- +-- VACUUM must not unlink a live posting tree page. +-- +-- When a page split is left unfinished, the new right half is reachable +-- through the sibling chain but has no downlink in its parent. VACUUM deletes +-- empty pages by walking downlinks, so it takes the incompletely split page +-- for the left sibling of the page it is deleting, and relinks the chain over +-- the top of the live right half. Every row indexed there becomes +-- unreachable. +-- +CREATE EXTENSION injection_points; +CREATE EXTENSION pageinspect; +-- Make injection points local to this process, for concurrency. +SELECT injection_points_set_local(); + injection_points_set_local +---------------------------- + +(1 row) + +-- One key with many rows, so that it gets a posting tree with several leaves. +CREATE TABLE gin_orphan (id int, a int[]) WITH (autovacuum_enabled = off); +INSERT INTO gin_orphan SELECT g, '{1}'::int[] FROM generate_series(1, 120000) g; +CREATE INDEX gin_orphan_idx ON gin_orphan USING gin (a) WITH (fastupdate = off); +-- Free heap space in the middle, so that the inserts below reuse those line +-- pointers and split a middle leaf. A split left unfinished at the rightmost +-- leaf would be harmless, as nothing is ever deleted to the right of it. +DELETE FROM gin_orphan WHERE id BETWEEN 40001 AND 80000; +VACUUM (INDEX_CLEANUP ON) gin_orphan; +-- Leave a split unfinished, as a crash or a cancelled insert would. +SELECT injection_points_attach('gin-leave-leaf-split-incomplete', 'error'); + injection_points_attach +------------------------- + +(1 row) + +DO $$ +DECLARE i int := 0; +BEGIN + LOOP + BEGIN + INSERT INTO gin_orphan VALUES (1000000 + i, '{1}'); + EXCEPTION WHEN others THEN + EXIT; + END; + i := i + 1; + IF i > 200000 THEN + RAISE 'no leaf split after % inserts', i; + END IF; + END LOOP; +END $$; +SELECT injection_points_detach('gin-leave-leaf-split-incomplete'); + injection_points_detach +------------------------- + +(1 row) + +-- The page that was split, its right half, and the page after that. +CREATE TEMP TABLE geom AS +SELECT b AS splitblk, + o.rightlink::int AS orphan, + (gin_page_opaque_info(get_raw_page('gin_orphan_idx', + o.rightlink::int))).rightlink::int AS victim + FROM generate_series(0, (pg_relation_size('gin_orphan_idx') / + current_setting('block_size')::int)::int - 1) b, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', b)) o + WHERE o.flags @> '{incomplete_split}'; +-- The tree has to take the shape the test needs, or the answers below could +-- come out right for uninteresting reasons. +SELECT count(*) = 1 AS one_unfinished_split, + bool_and(orph.flags @> '{data,leaf}' + AND NOT orph.flags @> '{deleted}') AS orphan_is_a_live_leaf, + bool_and(vic.rightlink <> 4294967295) AS victim_is_not_rightmost, + bool_and((SELECT count(*) > 0 + FROM gin_leafpage_items(get_raw_page('gin_orphan_idx', + g.orphan)) i, + LATERAL unnest(i.tids) t + WHERE t IN (SELECT ctid FROM gin_orphan))) AS orphan_holds_rows + FROM geom g, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.orphan)) orph, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.victim)) vic; + one_unfinished_split | orphan_is_a_live_leaf | victim_is_not_rightmost | orphan_holds_rows +----------------------+-----------------------+-------------------------+------------------- + t | t | t | t +(1 row) + +-- Empty exactly the page after the right half, so VACUUM deletes it. +DELETE FROM gin_orphan + WHERE ctid IN (SELECT t + FROM geom g, + gin_leafpage_items(get_raw_page('gin_orphan_idx', g.victim)) i, + LATERAL unnest(i.tids) t); +VACUUM (INDEX_CLEANUP ON) gin_orphan; +-- The chain must still reach the right half. +SELECT (gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.splitblk))).rightlink = g.orphan + AS chain_still_reaches_the_right_half + FROM geom g; + chain_still_reaches_the_right_half +------------------------------------ + t +(1 row) + +-- Every row left in the table has a = '{1}', so these two counts must agree. +SET enable_seqscan = off; +SELECT count(*) AS rows_found_by_the_index FROM gin_orphan WHERE a @> '{1}'; + rows_found_by_the_index +------------------------- + 81870 +(1 row) + +RESET enable_seqscan; +SELECT count(*) AS rows_in_the_table FROM gin_orphan; + rows_in_the_table +------------------- + 81870 +(1 row) + +DROP TABLE gin_orphan; +DROP EXTENSION pageinspect; +DROP EXTENSION injection_points; diff --git a/src/test/modules/gin/meson.build b/src/test/modules/gin/meson.build index ae2c1abe3..4c2cac17f 100644 --- a/src/test/modules/gin/meson.build +++ b/src/test/modules/gin/meson.build @@ -11,6 +11,7 @@ tests += { 'regress': { 'sql': [ 'gin_incomplete_splits', + 'gin_vacuum_orphan_page', ], }, } diff --git a/src/test/modules/gin/sql/gin_vacuum_orphan_page.sql b/src/test/modules/gin/sql/gin_vacuum_orphan_page.sql new file mode 100644 index 000000000..a2ee1ae5e --- /dev/null +++ b/src/test/modules/gin/sql/gin_vacuum_orphan_page.sql @@ -0,0 +1,96 @@ +-- +-- VACUUM must not unlink a live posting tree page. +-- +-- When a page split is left unfinished, the new right half is reachable +-- through the sibling chain but has no downlink in its parent. VACUUM deletes +-- empty pages by walking downlinks, so it takes the incompletely split page +-- for the left sibling of the page it is deleting, and relinks the chain over +-- the top of the live right half. Every row indexed there becomes +-- unreachable. +-- + +CREATE EXTENSION injection_points; +CREATE EXTENSION pageinspect; + +-- Make injection points local to this process, for concurrency. +SELECT injection_points_set_local(); + +-- One key with many rows, so that it gets a posting tree with several leaves. +CREATE TABLE gin_orphan (id int, a int[]) WITH (autovacuum_enabled = off); +INSERT INTO gin_orphan SELECT g, '{1}'::int[] FROM generate_series(1, 120000) g; +CREATE INDEX gin_orphan_idx ON gin_orphan USING gin (a) WITH (fastupdate = off); + +-- Free heap space in the middle, so that the inserts below reuse those line +-- pointers and split a middle leaf. A split left unfinished at the rightmost +-- leaf would be harmless, as nothing is ever deleted to the right of it. +DELETE FROM gin_orphan WHERE id BETWEEN 40001 AND 80000; +VACUUM (INDEX_CLEANUP ON) gin_orphan; + +-- Leave a split unfinished, as a crash or a cancelled insert would. +SELECT injection_points_attach('gin-leave-leaf-split-incomplete', 'error'); +DO $$ +DECLARE i int := 0; +BEGIN + LOOP + BEGIN + INSERT INTO gin_orphan VALUES (1000000 + i, '{1}'); + EXCEPTION WHEN others THEN + EXIT; + END; + i := i + 1; + IF i > 200000 THEN + RAISE 'no leaf split after % inserts', i; + END IF; + END LOOP; +END $$; +SELECT injection_points_detach('gin-leave-leaf-split-incomplete'); + +-- The page that was split, its right half, and the page after that. +CREATE TEMP TABLE geom AS +SELECT b AS splitblk, + o.rightlink::int AS orphan, + (gin_page_opaque_info(get_raw_page('gin_orphan_idx', + o.rightlink::int))).rightlink::int AS victim + FROM generate_series(0, (pg_relation_size('gin_orphan_idx') / + current_setting('block_size')::int)::int - 1) b, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', b)) o + WHERE o.flags @> '{incomplete_split}'; + +-- The tree has to take the shape the test needs, or the answers below could +-- come out right for uninteresting reasons. +SELECT count(*) = 1 AS one_unfinished_split, + bool_and(orph.flags @> '{data,leaf}' + AND NOT orph.flags @> '{deleted}') AS orphan_is_a_live_leaf, + bool_and(vic.rightlink <> 4294967295) AS victim_is_not_rightmost, + bool_and((SELECT count(*) > 0 + FROM gin_leafpage_items(get_raw_page('gin_orphan_idx', + g.orphan)) i, + LATERAL unnest(i.tids) t + WHERE t IN (SELECT ctid FROM gin_orphan))) AS orphan_holds_rows + FROM geom g, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.orphan)) orph, + LATERAL gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.victim)) vic; + +-- Empty exactly the page after the right half, so VACUUM deletes it. +DELETE FROM gin_orphan + WHERE ctid IN (SELECT t + FROM geom g, + gin_leafpage_items(get_raw_page('gin_orphan_idx', g.victim)) i, + LATERAL unnest(i.tids) t); +VACUUM (INDEX_CLEANUP ON) gin_orphan; + +-- The chain must still reach the right half. +SELECT (gin_page_opaque_info(get_raw_page('gin_orphan_idx', g.splitblk))).rightlink = g.orphan + AS chain_still_reaches_the_right_half + FROM geom g; + +-- Every row left in the table has a = '{1}', so these two counts must agree. +SET enable_seqscan = off; +SELECT count(*) AS rows_found_by_the_index FROM gin_orphan WHERE a @> '{1}'; +RESET enable_seqscan; + +SELECT count(*) AS rows_in_the_table FROM gin_orphan; + +DROP TABLE gin_orphan; +DROP EXTENSION pageinspect; +DROP EXTENSION injection_points; -- 2.53.0