6649594: Intermittent IOExceptions during dynamic attach on linux and solaris

[Daniel D. Daugherty]

On 7/14/2010 8:38 AM, Alan Bateman wrote:
>
> This is a long standing bug in the attach mechanism on Solaris and 
> Linux that causes intermittent failures for some tests. The bug is 
> that, on first attempt to attach, the client can observe the file to 
> which the door descriptor is attached or the socket file before 
> initialization is complete. The proposed fix removes the timing issue 
> by renaming the file into place after initialization. Some minor 
> clean-up while I was in the area. The webrev is here:
>  http://cr.openjdk.java.net/~alanb/6649594/webrev/
>
> Dan or Karen - one thing I wasn't sure about is whether to emit a 
> warning when the attach listener can't start or should warnings be 
> debug_only?
>
> -Alan.

src/os/linux/vm/attachListener_linux.cpp
    The addition of UNIX_PATH_MAX means that the size of the buffer
    on line 61: static char _path[PATH_MAX+1]
    is now or can be potentially out of sync with the size of the
    buffer on line 170 char path[UNIX_PATH_MAX]

    If UNIX_PATH_MAX happens to be greater than PATH_MAX, and the
    path to the socket is actually longer than PATH_MAX, then we'll
    store a short path in the buffer on line 61.

    Is there some reason not to use the new UNIX_PATH_MAX in both
    locations?

src/os/solaris/vm/attachListener_solaris.cpp
    The return values from snprintf() aren't checked here for potential
    overflow conditions. It seems to me that the Solaris logic could
    benefit from the same sanity checks that you make in the Linux
    version.

I would keep the warnings debug only. I see you have a debug only
warning in the Solaris code, but no warning in the Linux code. Or
were you talking about a different warning?

Dan