Re: [PATCH] s390/vfio_ccw: Fix target addresses of TIC CCWs

[Heiko Carstens <hca@linux.ibm.com>]

On Fri, Jun 28, 2024 at 09:31:56AM -0400, Eric Farman wrote:
> > > dma32_to_u32(ccw->cda) - ccw_head;
> > > -			ccw->cda = u32_to_dma32(cda);
> > > +			/* Calculate offset of TIC target */
> > > +			cda = dma32_to_u32(ccw->cda) - ccw_head;
> > > +			ccw->cda = virt_to_dma32(iter->ch_ccw) +
> > > cda;
> > 
> > I would suggest to rename cda to "offset", since that reflects what
> > it is
> > now. Also this code needs to take care of type checking, which will
> > fail now
> > due to dma32_t type (try "make C=1 drivers/s390/cio/vfio_ccw_cp.o).

...

> I was poking at that code yesterday because it seemed suspect, but as I
> wasn't getting an explicit failure (versus the CPC generated by hw), I
> opted to leave it for now. I agree they should both be fixed up.

...

> > I guess
> > you could add this hunk to your patch:
> > 
> > @@ -915,7 +915,7 @@ void cp_update_scsw(struct channel_program *cp,
> > union scsw *scsw)
> >  	 * in the ioctl directly. Path status changes etc.
> >  	 */
> >  	list_for_each_entry(chain, &cp->ccwchain_list, next) {
> > -		ccw_head = (u32)(u64)chain->ch_ccw;
> > +		ccw_head = (__force u32)virt_to_dma32(chain-
> > >ch_ccw);
> >  		/*
> >  		 * On successful execution, cpa points just beyond
> > the end
> >  		 * of the chain.

...

> > Furthermore it looks to me like the ch_iova member of struct ccwchain
> > should
> > get a dma32_t type instead of u64. The same applies to quite a few
> > variables
> > to the code. 
> 
> Agreed. I started this some time back after the IDAW code got reworked,
> but have been sidetracked. The problem with ch_iova is more apparent
> after the dma32 stuff.
> 
> > I could give this a try, but I think it would be better if
> > somebody who knows what he is doing would address this :)
> 
> I'll finish them up. But v2 will have to wait until after my holiday.
> Thanks for reminding me of the typechecking!

I hope you didn't get me wrong: from my point of view we want one or
two small patches (the above hunks), which fix the bugs, if you
agree.

And then address the type checking stuff at a later point in time.

(btw: your mailer adds lot's of extra line wraps)