[
https://issues.apache.org/jira/browse/HDFS-15641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17219003#comment-17219003
]
Hongbing Wang commented on HDFS-15641:
--------------------------------------
Thanks [~ferhui] for your reply. I will explain in two steps.
（a）*The occurrence of deadlock*: see figure below, and the corresponding jstack
is [^jstack.log]
!deadlock.png|width=973,height=214!
Related locks: `instance of BlockPoolManager` and `read-write lock in
BPOfferService`.
（b）*The fix I proposed:* In [^HDFS-15641.002.patch], I made 3 changes：
# `+BPOfferService.java+`: I just injected a test error to delay 1s. This
only takes effect in test and does not affect the production env. Both threads
will wait a short while after acquiring their respective locks.
# `+BPServiceActor.java+`: This is my change to fix the bug. Ensure that the
time to start `bpThread` is after the read lock is completed.
# `+TestRefreshNamenodesFailure.java+`: just test.
Merge the above 1 and 3 can reproduce the deadlock. And merge 1, 2 and 3 can
fix this deadlock.
The process after fixed is as follows:
!deadlock_fixed.png|width=1027,height=222!
Thanks again !
> DataNode could meet deadlock if invoke refreshNameNode
> ------------------------------------------------------
>
> Key: HDFS-15641
> URL: https://issues.apache.org/jira/browse/HDFS-15641
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 3.2.0
> Reporter: Hongbing Wang
> Assignee: Hongbing Wang
> Priority: Critical
> Attachments: HDFS-15641.001.patch, HDFS-15641.002.patch,
> deadlock.png, deadlock_fixed.png, jstack.log
>
>
> DataNode could meet deadlock when invoke `hdfs dfsadmin -refreshNamenodes
> hostname:50020` to register a new namespace in federation env.
> The jstack is shown in jstack.log
> The specific process is shown in Figure deadlock.png
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
