[
https://issues.apache.org/jira/browse/THRIFT-3038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15382679#comment-15382679
]
ASF GitHub Bot commented on THRIFT-3038:
----------------------------------------
Github user ben-craig commented on the issue:
https://github.com/apache/thrift/pull/981
The default op=, and even the default conversion operators from atomic<T>
to T are fine.
@nsuke I would prefer we keep the access to TThreadPoolServer timeout
variables as seq_cst. The setter operations shouldn't ever be in performance
critical code, and the 'get' side of things is happening in an already slow
area. On x86, the getters will still turn into regular 'mov' operations. On
ARM, there will be a couple of memory barriers generated, but those shouldn't
be a huge deal.
> Use of volatile in cpp library
> ------------------------------
>
> Key: THRIFT-3038
> URL: https://issues.apache.org/jira/browse/THRIFT-3038
> Project: Thrift
> Issue Type: Bug
> Components: C++ - Library
> Affects Versions: 0.9.2
> Reporter: Adriaan Schmidt
>
> In the cpp library there are several member variables which are declared
> volatile, I believe with the intention of providing some sort of
> thread-safety.
> While volatile can be used in this way in Java and C#, in C++ it cannot! It
> does not provide any guarantees with regard to instruction (re-)ordering,
> i.e. there are no implied memory barriers like you would get by using
> explicit locking or atomic variables.
> This means that all uses of volatile should be examined, the volatile
> qualifier should be removed and replaced by proper synchronization.
> The affected member variables are:
> # NoStarveReadWriteMutex::writerWaiting_
> Unprotected read access in acquireRead(). Data race can be seen by running
> the unit test with Helgrind.
> # (already fixed) TFileTransport::forceFlush_
> Always accessed while holding mutex_. In this case, the volatile can just be
> removed.
> # TFileTransport::closing_
> Sometimes accessed while holding mutex_ (in combination with the notEmpty_
> Monitor),
> but, e.g., enqueueEvent reads closing_ without any synchronization.
> # (already fixed) TThreadPoolServer::stop_, TThreadedServer::stop_
> Accessed (read and written) without synchronization. These would probably be
> fine using an atomic data type. Or, use explicit locking or signaling.
> # (already fixed) TThreadPoolServer::timeout_,
> TThreadPoolServer::taskExpiration_
> Should probably use a lock.
> # Mutex.cpp has mutexProfilingCounter as static variable. This probably
> doesn’t break anything, but still the volatile serves no real purpose.
> While some of the fixes are probably simple, in general I think someone with
> better knowledge of the code should have a look at this.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
