[jira] [Resolved] (OAK-12450) Concurrency and memory-retention issues in ThreadSpecificSegmentBufferWriterPool

["Nuno Santos (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/OAK-12450?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Nuno Santos resolved OAK-12450.
-------------------------------
    Fix Version/s: 2.8.0
       Resolution: Fixed

> Concurrency and memory-retention issues in ThreadSpecificSegmentBufferWriterPool
> --------------------------------------------------------------------------------
>
>                 Key: OAK-12450
>                 URL: https://issues.apache.org/jira/browse/OAK-12450
>             Project: Jackrabbit Oak
>          Issue Type: Bug
>          Components: segment-tar
>            Reporter: Lucas Weitzendorf
>            Assignee: Nuno Santos
>            Priority: Major
>             Fix For: 2.8.0
>
>
> {{ThreadSpecificSegmentBufferWriterPool}} has three problems:
> * *Read lock leak:* {{execute()}} gets the per-thread writer before entering {{try/finally}}. If creating the writer throws, the read lock is never released, and a later {{flush()}} (which takes the write lock) blocks forever.
> * *Data race on {{writerId}}:* {{writerId}} is a plain {{short}} incremented with {{++}} inside {{computeIfAbsent}}'s mapping function. That function can run concurrently for different keys, so updates can be lost and writer IDs can be duplicated.
> * *Dead threads kept alive:* the pool is keyed by {{Thread}}, so it holds terminated threads and their {{ThreadLocal}}s until the next {{flush()}}.
> Fix: get the writer inside {{try/finally}}, use an {{AtomicInteger}} with a bounded {{getAndUpdate}} for writer IDs, and key the pool by thread ID instead of {{Thread}}. Tests are added in {{SegmentBufferWriterPoolTest}}.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)