[ros-dev] Re: Race in rpc__cn_network_receiver()?

[Luke Kenneth Casson Leighton]

On Thu, Apr 20, 2006 at 03:04:21PM +1000, Luke Howard wrote:
 
 luke, you are a star: thank you for helping out with this.

 just for people's information (in case it's not been mentioned here - i
 must admit i forgot to notify opendce(a)opengroup.org!)

 about 6-8 months ago i did a coding-spree on freedce, with a view to
 achieving two, maybe three, major things:

 1) porting to win32 (if you use pthreads-win32)

 2) using posix final draft (7?) threads not posix draft 4.

 3) proving to the people on wine and reactos that they have
    a hell of a lot of work to do (measured in man-decades)
    by not using the open group's DCE/RPC reference
    implementation.
 
 here's the gotchas that i encountered, and had to give up:

 - udp worked, but tcp did not

 - reactos had some problems at the time with the code behind
   ipconfig.exe (NetTransportEnum i think it was)

 - pthreads-win32 was too strictly posix compliant (!)
   and there is a cancellation exception or something which
   makes it fail in the dcethreads exception library tests,
   where linux does not fail (!)
   
   maybe this is related to what lukeh is talking about.
   *shrug*.

 anyway, the upshot is that freedce/latest-cvs also doesn't work
 on TCP, and thank you so much people for helping track down why.

 it's only been ten years so far: hopefully some time within the next
 decade, freedce will be used for some real MSRPC-related free
 software projects.

 l


> >  In
> > rpc__cn_network_receiver(), rpc__socket_close() is called without an
> > exception handler in place. According to IEEE Std 1003.1 (2004), close()
> > is a cancellation point, see:
> >
> > http://www.opengroup.org/onlinepubs/009695399/functions/xsh_chap02_09.html
> >
> > Given this, a receiver thread could receive a cancel whilst closing the
> > socket, which will cause to entire program to abort. (We have seen this
> > occur when the RPC server is under load.)
> >
> > A suggested fix is to move the call to rpc__socket_close() inside the
> > TRY/CATCH_ALL block that removes any pending cancels. 
>  
>  It seems like the real fix is to disable cancellation within
>  rpc__socket_close(), because in other places the OSF code appears to
>  assume that this is not a cancellation point.
>  
>  Ditto for open(), socket(), etc (but not I/O routines that may block).
>  
>  I should note that the Linux dcethreads library does actually disable
>  cancellation for these calls, however in our own DCE thread library
>  we opted not to override any system calls to avoid changing semantics
>  when used by non-DCE code.
>  
>  -- Luke
>  
>  --
>   
-- 
--
lkcl.net - mad free software computer person, visionary and poet.
--