hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matt Foley (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-1790) review use of synchronized methods in FSNamesystem
Date Sat, 26 Mar 2011 19:47:05 GMT

    [ https://issues.apache.org/jira/browse/HDFS-1790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13011673#comment-13011673
] 

Matt Foley commented on HDFS-1790:
----------------------------------

SafeModeInfo contains several numeric variables which are overloaded with "in manual" or "in
extension" or "not actually in safemode" semantics based on special values -- instead of just
having a "state" variable (which I'm now providing).  Consequently, these variables may need
to be read/written in a synchronized way.  So it isn't unreasonable to synchronize them on
the SafeModeInfo instance.  But it seems to me that it is being done inconsistently, as described
above.

Do I understand you correctly, that you agree incVolumeFailure(), isInSafeMode() and isPopulatingReplQueues(),
should use the read/write lock rather than synchronizing on the FSNamesystem instance?  Thanks.

> review use of synchronized methods in FSNamesystem
> --------------------------------------------------
>
>                 Key: HDFS-1790
>                 URL: https://issues.apache.org/jira/browse/HDFS-1790
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: name-node
>    Affects Versions: 0.23.0
>            Reporter: Matt Foley
>             Fix For: 0.23.0
>
>
> While analyzing SafeMode semantics for HDFS-1726, I noticed the following usage of synchronized
methods in FSNamesystem that raised questions:
> 1. incVolumeFailure() is synchronized on the FSNamesystem instance (monitor lock), for
unclear reasons.  Is this a left-over from the conversion to read/write lock in FSNamesystem?
> 2. isInSafeMode() and isPopulatingReplQueues() are also synchronized on the FSNamesystem
instance (monitor lock), but the SafeModeInfo methods they call (safeMode.isInSafeMode() and
safeMode.isPopulatingReplQueues() are synchronized on the SafeModeInfo instance.  Is synchronizing
on the FSNamesystem instance necessary?  What does it add?  If it is necessary, should it
be using the read/write lock?
> 3. In SafeModeInfo, these methods are synchronized on the SafeModeInfo instance:
> - isOn
> - isPopulatingReplQueues
> - leave
> - initialize
> - canInitializeReplQueues
> - canLeave
> - setBlock
> - incrementSafeBlockCount
> - decrementSafeBlockCount
> - setManual
> but these are not:
> - enter
> - needEnter
> - checkMode
> - isManual
> - isConsistent
> Regarding these:
> - isOn() asserts isConsistent(), but is otherwise an atomic read operation.
> - isPopulatingReplQueues() is an atomic read.  Does it need synchronization?
> - leave() is complex, but shouldn't it be synchronized with enter(), which is also a
write operation?  Yet enter() is unsynchronized.
> - initializeReplQueues() calls blockManager.processMisReplicatedBlocks(), which can take
minutes to run.  By synchronizing on the SafeModeInfo instance, it prevents essentially all
of the other safeMode methods from running for the duration!  Is this desirable or needed?
> - canInitializeReplQueues() is a read-only operation, although it does compare two read
values.  needEnter() compares four read values, and is unsynchronized.  Does either need synchronization?
> - canLeave() is a compound operation, it's good that it is synchronized.
> - checkMode() is a big compound read/write operation.  Doesn't it need synchronization
if the other methods do?
> and so on.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message