[
https://issues.apache.org/jira/browse/HBASE-17872?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15959250#comment-15959250
]
stack commented on HBASE-17872:
-------------------------------
Seems a pity making runtime pay the price of test-time convenience. Volatile
reads are way more expensive than a local cache read. The UNSAFE_AVAIL check is
called frequently.
Make UNSAFE_AVAIL a final static?
Then it becomes how to change a static final setting at test time. Reflection?
A delegating class that makes use of a subclass of ByteBufferUtils ? Could
move the test clutter of detectAvailabilityOfUnsafe and disableUnsafe out of
this vital class and into this new test class.
> The MSLABImpl generates the invaild cells when unsafe is not availble
> ---------------------------------------------------------------------
>
> Key: HBASE-17872
> URL: https://issues.apache.org/jira/browse/HBASE-17872
> Project: HBase
> Issue Type: Bug
> Reporter: Chia-Ping Tsai
> Assignee: Chia-Ping Tsai
> Priority: Critical
> Fix For: 2.0.0
>
> Attachments: HBASE-17872.v0.patch, HBASE-17872.v1.patch,
> HBASE-17872.v2.patch
>
>
> We will get the wrong position of buffer in multithreaded environment, so the
> method makes the invalid cell in MSLAB.
> {noformat}
> public static int copyFromBufferToBuffer(ByteBuffer in, ByteBuffer out, int
> sourceOffset,
> int destinationOffset, int length) {
> if (in.hasArray() && out.hasArray()) {
> // ...
> } else if (UNSAFE_AVAIL) {
> // ...
> } else {
> int outOldPos = out.position();
> out.position(destinationOffset);
> ByteBuffer inDup = in.duplicate();
> inDup.position(sourceOffset).limit(sourceOffset + length);
> out.put(inDup);
> out.position(outOldPos);
> }
> return destinationOffset + length;
> }
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
