Re: [PATCH 20/22] iommu/amd: Implement IO page-fault handler

[Joerg Roedel <joro@8bytes.org>]

On Wed, Dec 14, 2011 at 11:13:00AM -0500, Jerome Glisse wrote:
> On Mon, Dec 05, 2011 at 02:34:35PM +0100, Joerg Roedel wrote:
> > +	npages = get_user_pages(fault->state->task, fault->state->mm,
> > +				fault->address, 1, write, 0, &page, NULL);
> > +
> > +	if (npages == 1)
> > +		put_page(page);
> > +	else
> > +		set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
> 
> I might be missing something on how mm work in linux but can't the page
> in a vma change anytime (like being migrated to a node or being evicted
> ...) I guess my question is when a fault happen is it because the task
> page table have an invalid entry ?

Yes, but this is not a problem. When the page-fault handler signals
completion of the page-fault back to the device it will start to re-walk
the page-table. If the page-walker detects that the page is not there it
will just signal the fault again. MMU-Notifiers take care of the
TLB management.

> I was under the impression that their was a page table associated with
> each iommu client. Thus that the task page table was never directly use.
> But i haven't carrefully read all the patches.

The IOMMUv2 can handle x86 long-mode page-tables (including
accessed/dirty bits). So it is easy to just re-use the task page-tables
in the IOMMUv2.


Regards,

	Joerg