Re: [patch] Staging: hv: Fix vmbus load hang caused by wrong data packing

[Greg KH <gregkh@suse.de>]

On Mon, Oct 12, 2009 at 08:44:06PM +0000, Haiyang Zhang wrote:
> > Can you figure out why?  What is the output of gcc for both ways?
> There are no build errors in either ways, I think it may be a bug in
> gcc handing __attribute__((packed)).

Of course there is not a build error.  I mean what is the difference in
the assembly produced?  Why is the __attribute__ version not the same

> > Can you show what is fixed by this change?
> Before the fix:
> The command "insmod hv_vmbus.ko" hangs. 
> Root cause: The message data sent from Linux guest to HyperV host were
> not correctly packed and not recognized by HyperV host. So, the host
> doesn't acknowledge that the vmbus channel is set up. Then, the guest
> keeps on waiting for the event and hangs the insmod command.
> 
> After the fix:
> The command "insmod hv_vmbus.ko" completes successfully with vmbus
> module loaded.

That's not showing me why the above fixed the problem.

> > Also note that #pragma packed is not supported by older versions of gcc,
> > so I don't think that it would work at all on some compiler versions
> > that are still legal to use for the kernel.  But I'm not quite sure when
> > it was added, so I might be wrong.
> This pragma was used in older distro, such as SLES10 (kernel 2.6.16)
> without any problem.

What compiler version was that?  Have you tried it on the lowest
required version of gcc?

> >From your newer email:
> > I also note that you are packing more structures here with this patch than the code before the patch had.  Could that be the 
> > issue here?  The additional structures are ones that should be packed instead?
> No, the extra packed structures are not required for this bug fix. Just to make sure any future struct related to message is packed, so I put the pack() around the entire file.

Please wrap your email lines properly...

Your patch changes structures that it is needed for.

We need to find the root cause of _why_ your patch fixes the problem here,
that is what I am trying to get you to do.

thanks,

greg k-h