Re: [RFC PATCH 2/3] vfio-pci: Allow to mmap sub-page MMIO BARs if all MMIO BARs are page aligned

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 2015-12-17 at 18:26 +0800, yongji xie wrote:
> 
> On 2015/12/17 4:04, Alex Williamson wrote:
> > On Fri, 2015-12-11 at 16:53 +0800, Yongji Xie wrote:
> > > Current vfio-pci implementation disallows to mmap
> > > sub-page(size < PAGE_SIZE) MMIO BARs because these BARs' mmio
> > > page
> > > may be shared with other BARs.
> > > 
> > > But we should allow to mmap these sub-page MMIO BARs if all MMIO
> > > BARs
> > > are page aligned which leads the BARs' mmio page would not be
> > > shared
> > > with other BARs.
> > > 
> > > This patch adds support for this case and we also add a
> > > VFIO_DEVICE_FLAGS_PCI_PAGE_ALIGNED flag to notify userspace that
> > > platform supports all MMIO BARs to be page aligned.
> > > 
> > > Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com>
> > > ---
> > >   drivers/vfio/pci/vfio_pci.c         |   10 +++++++++-
> > >   drivers/vfio/pci/vfio_pci_private.h |    5 +++++
> > >   include/uapi/linux/vfio.h           |    2 ++
> > >   3 files changed, 16 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/vfio/pci/vfio_pci.c
> > > b/drivers/vfio/pci/vfio_pci.c
> > > index 32b88bd..dbcad99 100644
> > > --- a/drivers/vfio/pci/vfio_pci.c
> > > +++ b/drivers/vfio/pci/vfio_pci.c
> > > @@ -443,6 +443,9 @@ static long vfio_pci_ioctl(void *device_data,
> > >   		if (vdev->reset_works)
> > >   			info.flags |= VFIO_DEVICE_FLAGS_RESET;
> > >   
> > > +		if (vfio_pci_bar_page_aligned())
> > > +			info.flags |=
> > > VFIO_DEVICE_FLAGS_PCI_PAGE_ALIGNED;
> > > +
> > >   		info.num_regions = VFIO_PCI_NUM_REGIONS;
> > >   		info.num_irqs = VFIO_PCI_NUM_IRQS;
> > >   
> > > @@ -479,7 +482,8 @@ static long vfio_pci_ioctl(void *device_data,
> > >   				     VFIO_REGION_INFO_FLAG_WRIT
> > > E;
> > >   			if (IS_ENABLED(CONFIG_VFIO_PCI_MMAP) &&
> > >   			    pci_resource_flags(pdev,
> > > info.index) &
> > > -			    IORESOURCE_MEM && info.size >=
> > > PAGE_SIZE)
> > > +			    IORESOURCE_MEM && (info.size >=
> > > PAGE_SIZE ||
> > > +			    vfio_pci_bar_page_aligned()))
> > >   				info.flags |=
> > > VFIO_REGION_INFO_FLAG_MMAP;
> > >   			break;
> > >   		case VFIO_PCI_ROM_REGION_INDEX:
> > > @@ -855,6 +859,10 @@ static int vfio_pci_mmap(void *device_data,
> > > struct vm_area_struct *vma)
> > >   		return -EINVAL;
> > >   
> > >   	phys_len = pci_resource_len(pdev, index);
> > > +
> > > +	if (vfio_pci_bar_page_aligned())
> > > +		phys_len = PAGE_ALIGN(phys_len);
> > > +
> > >   	req_len = vma->vm_end - vma->vm_start;
> > >   	pgoff = vma->vm_pgoff &
> > >   		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) -
> > > 1);
> > > diff --git a/drivers/vfio/pci/vfio_pci_private.h
> > > b/drivers/vfio/pci/vfio_pci_private.h
> > > index 0e7394f..319352a 100644
> > > --- a/drivers/vfio/pci/vfio_pci_private.h
> > > +++ b/drivers/vfio/pci/vfio_pci_private.h
> > > @@ -69,6 +69,11 @@ struct vfio_pci_device {
> > >   #define is_irq_none(vdev) (!(is_intx(vdev) || is_msi(vdev) ||
> > > is_msix(vdev)))
> > >   #define irq_is(vdev, type) (vdev->irq_type == type)
> > >   
> > > +static inline bool vfio_pci_bar_page_aligned(void)
> > > +{
> > > +	return IS_ENABLED(CONFIG_PPC64);
> > > +}
> > I really dislike this.  This is a problem for any architecture that
> > runs on larger pages, and even an annoyance on 4k hosts.  Why are
> > we
> > only solving it for PPC64?
> Yes, I know it's a problem for other architectures. But I'm not sure
> if 
> other archs prefer
> to enforce the alignment of all BARs to be at least PAGE_SIZE which 
> would result in
> some waste of address space.
> 
> So I just propose a prototype and add PPC64 support here. And other 
> archs could decide
> to use it or not by themselves.
> > Can't we do something similar in the core PCI code and detect it?
> So you mean we can do it like this:
> 
> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h
> index d390fc1..f46c04d 100644
> --- a/drivers/pci/pci.h
> +++ b/drivers/pci/pci.h
> @@ -320,6 +320,11 @@ static inline resource_size_t 
> pci_resource_alignment(struct pci_dev *dev,
>          return resource_alignment(res);
>   }
> 
> +static inline bool pci_bar_page_aligned(void)
> +{
> +       return IS_ENABLED(CONFIG_PPC64);
> +}
> +
>   void pci_enable_acs(struct pci_dev *dev);
> 
>   struct pci_dev_reset_methods {
> 
> or add a config option to indicate that PCI MMIO BARs should be page 
> aligned? 

Yes, I'm thinking of a boot commandline option, maybe one that PPC64
can default to enabled if it chooses to.  The problem is not unique to
PPC64 and the solution should not be unique either.  I don't want to
need to revisit this for ARM, which we know is going to be similarly
afflicted.  Thanks,

Alex