[ 
https://issues.apache.org/jira/browse/HBASE-22480?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16849298#comment-16849298
 ] 

Zheng Hu commented on HBASE-22480:
----------------------------------

There're several place where we did not consider the release or release 
incorrectly I think: 
1.  in seekBefore method ( see 
[Line-882|https://github.com/apache/hbase/blob/8e47c8e73943b9595b2a8bfea131b106a3947fbd/hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderImpl.java#L882]),
  if the previousBlockOffset is -1,  we return -1 without returning the 
seekToBlock back to blockcache,  it's a memory leak issue; 
2. in updateCurrentBlock(see 
[L1199|https://github.com/apache/hbase/blob/8e47c8e73943b9595b2a8bfea131b106a3947fbd/hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderImpl.java#L1199])
 method (Both HFileScannerImpl and EncodedScanner),  there's possible that we 
did not make the curBlock refer to the newBlock , such as some exception happen 
or the newBlock has the same offset but another instance, etc. 

So I think a generic method to *release the newBlock but except the curBlock* 
will be a better abstraction to handle all those places.  I also handled the 
similar issues In my PR for branch HBASE-21879 
[PR-257|https://github.com/apache/hbase/pull/257/files#diff-f14c59a7641b9c63cf302f41af83fa35].
 

> Get block from BlockCache once and return this block to BlockCache twice make 
> ref count error.
> ----------------------------------------------------------------------------------------------
>
>                 Key: HBASE-22480
>                 URL: https://issues.apache.org/jira/browse/HBASE-22480
>             Project: HBase
>          Issue Type: Sub-task
>            Reporter: binlijin
>            Assignee: binlijin
>            Priority: Major
>         Attachments: HBASE-22480-master-v1.patch, HBASE-22480-master-v2.patch
>
>
> After debugging HBASE-22433, i find the problem it is that we get a block 
> from BucketCache once and return this block to BucketCache twice and make the 
> ref count error, sometimes the refCount can be negative.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to