From 98f52d6bb5c0464465ab75f48ae3dc5d9a18c99f Mon Sep 17 00:00:00 2001 From: Oleg Tkachenko Date: Fri, 16 Jan 2026 12:54:14 +0100 Subject: [PATCH] Clamp truncation_block_length to RELSEG_SIZE when sending incremental files - Fix bug where truncation_block_length could exceed RELSEG_SIZE, leading to restore failures. - Add TAP test to verify that a relation truncated after the last base backup can be safely included in an incremental backup. While fixing the issue with the truncation of the main fork, we discovered another bug in the incremental backup. When WAL summaries are processed, the incremental backup code incorrectly assumes that truncation of the main fork also applies to the VM fork. As a result, after restore, the VM fork file can end up with an incorrect size. At the time of writing, it was decided to postpone fixing the VM fork issue and proceed only with the fix for the main fork truncation. This patch follows guidance from Robert Haas. --- src/backend/backup/basebackup_incremental.c | 14 ++ src/bin/pg_combinebackup/meson.build | 1 + ...011_incremental_backup_truncation_block.pl | 131 ++++++++++++++++++ 3 files changed, 146 insertions(+) create mode 100644 src/bin/pg_combinebackup/t/011_incremental_backup_truncation_block.pl diff --git a/src/backend/backup/basebackup_incremental.c b/src/backend/backup/basebackup_incremental.c index 891af7cad8e..77dce24ad38 100644 --- a/src/backend/backup/basebackup_incremental.c +++ b/src/backend/backup/basebackup_incremental.c @@ -850,8 +850,22 @@ GetFileBackupMethod(IncrementalBackupInfo *ib, const char *path, { unsigned relative_limit = limit_block - segno * RELSEG_SIZE; + /* + * We can't set a truncation_block_length in excess of the limit block + * number (relativized to the current segment). To do so would be to + * treat blocks from older backups as valid current contents even if + * they were subsequently truncated away. + */ if (*truncation_block_length < relative_limit) *truncation_block_length = relative_limit; + + /* + * We also can't set a truncation_block_length in excess of the + * segment size, since the reconstructed file can't be larger than + * that. + */ + if (*truncation_block_length > RELSEG_SIZE) + *truncation_block_length = RELSEG_SIZE; } /* Send it incrementally. */ diff --git a/src/bin/pg_combinebackup/meson.build b/src/bin/pg_combinebackup/meson.build index 44780927cd6..613f73a756b 100644 --- a/src/bin/pg_combinebackup/meson.build +++ b/src/bin/pg_combinebackup/meson.build @@ -38,6 +38,7 @@ tests += { 't/008_promote.pl', 't/009_no_full_file.pl', 't/010_hardlink.pl', + 't/011_incremental_backup_truncation_block.pl', ], } } diff --git a/src/bin/pg_combinebackup/t/011_incremental_backup_truncation_block.pl b/src/bin/pg_combinebackup/t/011_incremental_backup_truncation_block.pl new file mode 100644 index 00000000000..6c9e39e84e5 --- /dev/null +++ b/src/bin/pg_combinebackup/t/011_incremental_backup_truncation_block.pl @@ -0,0 +1,131 @@ +# Test case for a scenario where truncation_block is miscalculated +# during an incremental backup + +use strict; +use warnings FATAL => 'all'; +use PostgreSQL::Test::Cluster; +use PostgreSQL::Test::Utils; +use Test::More; + +# Initialize primary node +my $primary = PostgreSQL::Test::Cluster->new('primary'); +$primary->init(has_archiving => 1, allows_streaming => 1); +# Enable WAL summarization to support incremental backup +$primary->append_conf('postgresql.conf', 'summarize_wal = on'); +$primary->start; + +# Backup locations +my $backup_path = $primary->backup_dir; +my $full_backup = "$backup_path/full"; + +# While fixing the issue with the truncation of the main fork, +# we discovered another bug in the incremental backup. +# When WAL summaries are processed, the incremental backup code +# incorrectly assumes that truncation of the main fork also applies to +# the VM fork. As a result, after restore, the VM fork file can end up +# with an incorrect size. +# +# At the time of writing, it was decided to postpone fixing the VM fork +# issue and proceed only with the fix for the main fork truncation. +# +# !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! +# Some test platforms are constrained by disk space, so creating a +# 1GB VM fork file is not possible there. +# Other platforms configure the segment size to six blocks, +# which means that only on those platforms this test actually tests the +# problematic code path. +# !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! +# +# Therefore, the test is written to produce a relation that spans more +# than one segment but remains small on most systems. We first create +# enough pages to exceed a single segment, then remove rows to make +# truncation possible. +my $segment_size_blocks = 6; +my $block_size = $primary->safe_psql('postgres', + "SELECT current_setting('block_size')::int;"); + +# Two blocks will be located in the second segment, and one block +# will still stay there after truncation. +my $target_rows = int($segment_size_blocks + 2); +my $rows_after_truncation = int($target_rows - 1); + +# Create a test table. +# STORAGE PLAIN prevents compression and TOASTing of repetitive data, +# ensuring predictable row sizes. +$primary->safe_psql( + 'postgres', q{ + CREATE TABLE t ( + id int, + data text STORAGE PLAIN + ); +}); + +# The tuple size should be enough to prevent two tuples +# from being on the same page. +# Since the template string has a length of 32 bytes, +# it's enough to repeat it (block_size / (2*32)) times. +$primary->safe_psql( + 'postgres', + "INSERT INTO t + SELECT i, + repeat('0123456789ABCDEF0123456789ABCDEF', ($block_size / (2*32))) + FROM generate_series(1, $target_rows) i;" +); + +# This step is required because at this moment, +# tuples do not have hint bits set. +# Later (for example, soon after the base backup is created), a background +# process may set hint bits on many tuples and change many heap pages. +# Because of this, the WAL summary may show that too many pages were changed +# and create a full file copy instead of an incremental one, which makes the +# issue non-reproducible. +$primary->safe_psql('postgres', 'VACUUM t;'); + +# Verify that relation spans more than one physical storage segment +my $t_blocks = $primary->safe_psql('postgres', + "SELECT pg_relation_size('t') / current_setting('block_size')::int;"); +cmp_ok($t_blocks, '>', $segment_size_blocks, + 'Relation spans more than one physical segment.'); + +# Take a full base backup +$primary->backup('full'); + +# Delete rows at the logical end of the table. +# This creates removable empty pages at the tail +$primary->safe_psql('postgres', + "DELETE FROM t WHERE id > ($rows_after_truncation);"); + +# Although TRUNCATE is enabled by default, +# here it emphasizes the expected behavior of the operation. +$primary->safe_psql('postgres', 'VACUUM (TRUNCATE) t;'); + +# Verify that after VACUUM relation is truncated but still spans more than one +# physical storage segment. +$t_blocks = $primary->safe_psql('postgres', + "SELECT pg_relation_size('t') / current_setting('block_size')::int;"); +is($t_blocks, $rows_after_truncation, 'Relation has expected size.'); +cmp_ok($t_blocks, '>', $segment_size_blocks, + 'Relation spans more than one physical segment.'); + +# Take an incremental backup based on the full backup manifest +$primary->backup('incr', + backup_options => [ '--incremental', "$full_backup/backup_manifest" ]); + +# Combine full and incremental backups. +# This step must correctly handle truncated relation segments. +# Before the fix, this failed because the INCREMENTAL file header +# contained an incorrect truncation_block value. +my $restored = PostgreSQL::Test::Cluster->new('node2'); +$restored->init_from_backup($primary, 'incr', combine_with_prior => ['full']); +$restored->start(); + +# Check that the restored table contains the correct number of rows +my $restored_count = + $restored->safe_psql('postgres', "SELECT count(*) FROM t;"); +is($restored_count, $rows_after_truncation, + 'Restored backup has correct row count'); + +$primary->stop; +$restored->stop; + +done_testing(); -- 2.52.0