[jira] [Commented] (THRIFT-3224) Fix TNamedPipeServer unpredictable behavior on accept

["Ben Craig (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/THRIFT-3224?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14617344#comment-14617344 ] 

Ben Craig commented on THRIFT-3224:
-----------------------------------

This patch leaks pipes if the TPipe ctor throws.

There is a bug here, but this isn't the right place to fix it.

The pattern that needs to be followed is 'do everything that can fail first, then commit'.  The existing code messes that up in TPipe.cpp, TWaitableNamedPipeImpl ctor.  Pipe_ shouldn't be given ownership of the passed in pipe until after all the failing operations have completed (mainly buffer.resize() and beginAsyncRead() ).

> Fix TNamedPipeServer unpredictable behavior on accept
> -----------------------------------------------------
>
>                 Key: THRIFT-3224
>                 URL: https://issues.apache.org/jira/browse/THRIFT-3224
>             Project: Thrift
>          Issue Type: Bug
>          Components: C++ - Library
>    Affects Versions: 0.9.2
>         Environment: Windows
>            Reporter: Paweł Janicki
>            Assignee: Paweł Janicki
>            Priority: Critical
>              Labels: patch
>         Attachments: 0001-THRIFT-3224.-cpp-Fix-TNamedPipeServer-unpredictable-.patch
>
>   Original Estimate: 2h
>  Remaining Estimate: 2h
>
> Application bahavior utilizing TNamedPipeServer is unpredictable due misuse of TAutoHandle.
> Project uses TAutoHandle class, an analogy of std::unique_ptr, for managing WIN32 handles. The dangerous members of this concept are: the direct getter "HANDLE TAutoHandle::h" and release method "void __thiscall TAutoHandle::release()" 
> Below code citation introduces serous bug:
> {code:java}
> {
>     TAutoCrit lock(pipe_protect_);
>     GlobalOutput.printf("Client connected.");
>     shared_ptr<TPipe> client(new TPipe(Pipe_.h));
>     Pipe_.release();
> }
> {code}
> The getter is  used in TNamedPipeServer::acceptImpl() to pass internal  handle value to c-tor of TPipe and just after c-tion HANDLE__thiscall TAutoHandle::release() is called to release ownership. That means the TPipe object is expected to take ownership of the resource, but if TPipe c-tor throws the d-tor of TAutoHandle is called releasing the resource and the incomplete TPipe object does the same. Since now it is impossible to ensure that the second free of the handle value was not performed on a resource that was opened in meantime by other thread.
> I propose to solve the issue by ensuring the handle is not owned by two objects at a time.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)