Re: [Xen-devel] [PATCH] xen/swiotlb: Exchange to contiguous memory for map_sg hook

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Mon, Jan 07, 2013 at 07:17:33AM +0000, Xu, Dongxiao wrote:
> > -----Original Message-----
> > From: Jan Beulich [mailto:JBeulich@suse.com]
> > Sent: Thursday, December 20, 2012 4:56 PM
> > To: Xu, Dongxiao
> > Cc: xen-devel@lists.xen.org; Konrad Rzeszutek Wilk;
> > linux-kernel@vger.kernel.org
> > Subject: RE: [Xen-devel] [PATCH] xen/swiotlb: Exchange to contiguous memory
> > for map_sg hook
> > 
> > >>> On 20.12.12 at 02:23, "Xu, Dongxiao" <dongxiao.xu@intel.com> wrote:
> > > Sorry, maybe I am still not describing this issue clearly.
> > 
> > No, at least I understood you the way you re-describe below.
> > 
> > > Take the libata case as an example, the static DMA buffer locates
> > > (dev->link->ap->sector_buf , here we use Data Structure B in the graph) in
> > > the following structure:
> > >
> > > -------------------------------------Page boundary
> > > <Data Structure A>
> > > <Data Structure B>
> > > -------------------------------------Page boundary
> > > <Data Structure B (cross page)>
> > > <Data Structure C>
> > > -------------------------------------Page boundary
> > >
> > > Where Structure B is our DMA target.
> > >
> > > For Data Structure B, we didn't care about the simultaneous access, either
> > > lock or sync function will take care of it.
> > > What we are not sure is "read/write of A and C from other processor". As we
> > > will have memory copy for the pages, and at the same time, other CPU may
> > > access A/C.
> > 
> > The question is whether what libata does here is valid in the first
> > place - fill an SG list entry with something that crosses a page
> > boundary and is not a compound page.
> 
> Sorry for the late response about this thread.
> 
> To make sure I understand you correctly, so do you mean the correct fix should be applied to libata driver, and make sure it DMA from/to correct place (for example, some memory allocated by DMA API), but not such certain field in a static structure?

Or with baremetal swiotlb if the user booted with 'swiotlb=force'.

> 
> If we fix it in device driver side, then we may not need to modify the xen-swiotlb part.
> 
Correct. This is a bug in the device driver where it checks the contents of its buffer
_before_ doing an DMA unmap or DMA sync.