You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Sergey Korotkov (Jira)" <ji...@apache.org> on 2025/05/28 04:06:00 UTC
[jira] [Comment Edited] (IGNITE-17734) CorruptedFreeListException after upgrade to 2.13.0 and JDK17
[ https://issues.apache.org/jira/browse/IGNITE-17734?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17954476#comment-17954476 ]
Sergey Korotkov edited comment on IGNITE-17734 at 5/28/25 4:05 AM:
-------------------------------------------------------------------
It appears that the C2 JIT compiler optimizations cause this incorrect behavior (exclusively in case of ShenandoahGC and jdk17/jdk21).
In the _PagesList$CutTail.run()_ method, the reading and writing of nextId are reordered. So _tailId_ is always 0. If asserts are active, then checking _tailId_ for 0 prevents this reordering. So fix is to add _tailId_ check irrespectively on asserts presence:
!fix.png|width=800!
The effect is visible in the generated assembler code.
Note that 0x30 (PagesListNodeIO.NEXT_PAGE_ID_OFF) is exactly the offset in the page where the pointer to the next one in FreeList is stored.
Shenandoah GC and {*}no asserts{*}: [^shenandoah-CutTailRun.asm]
!shenandoah_and_no_asserts.png|width=800!
Shenandoah GC and {*}asserts{*}: [^shenandoah-with-asserts-CutTailRun.asm]
!shenandoah_and_asserts.png|width=800!
In the case G1GC and there are {*}no asserts{*}, the optimization is less aggressive and no reordering is observed: [^g1-CutTailRun.asm]
!G1_and_no_asserts.png|width=800!
***
Test shows that jdk11 is not affected for any of garbage collectors (G1GC, ShenandoahGC, ZGC are tested)
jdk17 and jdk21 with ShenandoahGC are affected.
***
Assembler code for cases above are produced with java version:
{noformat}
openjdk version "17.0.15" 2025-04-15
OpenJDK Runtime Environment (build 17.0.15+6-Ubuntu-0ubuntu124.04)
OpenJDK 64-Bit Server VM (build 17.0.15+6-Ubuntu-0ubuntu124.04, mixed mode, sharing)
{noformat}
was (Author: JIRAUSER279895):
It appears that the C2 JIT compiler optimizations cause this incorrect behavior (exclusively in case of ShenandoahGC and jdk17/jdk21).
In the _PagesList$CutTail.run()_ method, the reading and writing of nextId are reordered. So _tailId_ is always 0. If asserts are active, then checking _tailId_ for 0 prevents this reordering. So fix is to add _tailId_ check irrespectively on asserts presence:
!fix.png|width=800!
The effect is visible in the generated assembler code.
Note that 0x30 (PagesListNodeIO.NEXT_PAGE_ID_OFF) is exactly the offset in the page where the pointer to the next one in FreeList is stored.
Shenandoah GC and {*}no asserts{*}:
!shenandoah_and_no_asserts.png|width=800!
Shenandoah GC and {*}asserts{*}:
!shenandoah_and_asserts.png|width=800!
In the case G1GC and there are {*}no asserts{*}, the optimization is less aggressive and no reordering is observed:
!G1_and_no_asserts.png|width=800!
***
Test shows that jdk11 is not affected for any of garbage collectors (G1GC, ShenandoahGC, ZGC are tested)
jdk17 and jdk21 with ShenandoahGC are affected.
***
Full assembler code for cases above are attached for references. Produced on the below java version
{noformat}
openjdk version "17.0.15" 2025-04-15
OpenJDK Runtime Environment (build 17.0.15+6-Ubuntu-0ubuntu124.04)
OpenJDK 64-Bit Server VM (build 17.0.15+6-Ubuntu-0ubuntu124.04, mixed mode, sharing)
{noformat}
> CorruptedFreeListException after upgrade to 2.13.0 and JDK17
> ------------------------------------------------------------
>
> Key: IGNITE-17734
> URL: https://issues.apache.org/jira/browse/IGNITE-17734
> Project: Ignite
> Issue Type: Bug
> Components: cache
> Affects Versions: 2.12, 2.13, 2.14
> Environment: JDK17, IGNITE 2.13.0
> Reporter: Lopata
> Assignee: Sergey Korotkov
> Priority: Major
> Labels: ise
> Fix For: 2.18
>
> Attachments: FreeListTest.java, G1_and_no_asserts.png, fix.png, g1-CutTailRun.asm, image-2022-12-03-09-31-27-347.png, shenandoah-CutTailRun.asm, shenandoah-with-asserts-CutTailRun.asm, shenandoah_and_asserts.png, shenandoah_and_no_asserts.png, stacktrace.txt
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> After migration to Ignite version 2.13.0 from 2.12.0 we are getting following error:
> org.apache.ignite.internal.processors.cache.persistence.freelist.CorruptedFreeListException: Failed to insert data row
> caused by:
> Caused by: java.lang.IllegalStateException: Tail not found: 0
>
> Full stack trace is in attachment. Along with the Ignite version upgrade, we are updating Java from 11 to 17. There have been no other changes in our code or IGNITE configurations. The error occurs after a few hours of running and then occurs for every store.
> Can you help us to to solve this problem.
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)