Re: [PATCH v17 Kernel 6/7] vfio iommu: Adds flag to indicate dirty pages tracking capability support

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

On Wed, 1 Apr 2020 11:51:03 -0600
Alex Williamson <alex.williamson@redhat.com> wrote:

> On Wed, 1 Apr 2020 22:55:57 +0530
> Kirti Wankhede <kwankhede@nvidia.com> wrote:
> 
> > On 4/1/2020 12:45 AM, Alex Williamson wrote:  
> > > On Wed, 1 Apr 2020 00:38:49 +0530
> > > Kirti Wankhede <kwankhede@nvidia.com> wrote:
> > >     
> > >> On 3/31/2020 2:28 AM, Alex Williamson wrote:    
> > >>> On Mon, 30 Mar 2020 22:20:43 +0530
> > >>> Kirti Wankhede <kwankhede@nvidia.com> wrote:
> > >>>        
> > >>>> Flag VFIO_IOMMU_INFO_DIRTY_PGS in VFIO_IOMMU_GET_INFO indicates that driver
> > >>>> support dirty pages tracking.
> > >>>>
> > >>>> Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
> > >>>> Reviewed-by: Neo Jia <cjia@nvidia.com>
> > >>>> ---
> > >>>>    drivers/vfio/vfio_iommu_type1.c | 3 ++-
> > >>>>    include/uapi/linux/vfio.h       | 5 +++--
> > >>>>    2 files changed, 5 insertions(+), 3 deletions(-)
> > >>>>
> > >>>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> > >>>> index 266550bd7307..9fe12b425976 100644
> > >>>> --- a/drivers/vfio/vfio_iommu_type1.c
> > >>>> +++ b/drivers/vfio/vfio_iommu_type1.c
> > >>>> @@ -2390,7 +2390,8 @@ static long vfio_iommu_type1_ioctl(void *iommu_data,
> > >>>>    			info.cap_offset = 0; /* output, no-recopy necessary */
> > >>>>    		}
> > >>>>    
> > >>>> -		info.flags = VFIO_IOMMU_INFO_PGSIZES;
> > >>>> +		info.flags = VFIO_IOMMU_INFO_PGSIZES |
> > >>>> +			     VFIO_IOMMU_INFO_DIRTY_PGS;
> > >>>>    
> > >>>>    		info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
> > >>>>    
> > >>>> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
> > >>>> index e3cbf8b78623..0fe7c9a6f211 100644
> > >>>> --- a/include/uapi/linux/vfio.h
> > >>>> +++ b/include/uapi/linux/vfio.h
> > >>>> @@ -985,8 +985,9 @@ struct vfio_device_feature {
> > >>>>    struct vfio_iommu_type1_info {
> > >>>>    	__u32	argsz;
> > >>>>    	__u32	flags;
> > >>>> -#define VFIO_IOMMU_INFO_PGSIZES (1 << 0)	/* supported page sizes info */
> > >>>> -#define VFIO_IOMMU_INFO_CAPS	(1 << 1)	/* Info supports caps */
> > >>>> +#define VFIO_IOMMU_INFO_PGSIZES   (1 << 0) /* supported page sizes info */
> > >>>> +#define VFIO_IOMMU_INFO_CAPS      (1 << 1) /* Info supports caps */
> > >>>> +#define VFIO_IOMMU_INFO_DIRTY_PGS (1 << 2) /* supports dirty page tracking */
> > >>>>    	__u64	iova_pgsizes;	/* Bitmap of supported page sizes */
> > >>>>    	__u32   cap_offset;	/* Offset within info struct of first cap */
> > >>>>    };    
> > >>>
> > >>>
> > >>> As I just mentioned in my reply to Yan, I'm wondering if
> > >>> VFIO_CHECK_EXTENSION would be a better way to expose this.  The
> > >>> difference is relatively trivial, but currently the only flag
> > >>> set by VFIO_IOMMU_GET_INFO is to indicate the presence of a field in
> > >>> the returned structure.  I think this is largely true of other INFO
> > >>> ioctls within vfio as well and we're already using the
> > >>> VFIO_CHECK_EXTENSION ioctl to check supported IOMMU models, and IOMMU
> > >>> cache coherency.  We'd simply need to define a VFIO_DIRTY_PGS_IOMMU
> > >>> value (9) and return 1 for that case.  Then when we enable support for
> > >>> dirt pages that can span multiple mappings, we can add a v2 extensions,
> > >>> or "MULTI" variant of this extension, since it should be backwards
> > >>> compatible.
> > >>>
> > >>> The v2/multi version will again require that the user provide a zero'd
> > >>> bitmap, but I don't think that should be a problem as part of the
> > >>> definition of that version (we won't know if the user is using v1 or
> > >>> v2, but a v1 user should only retrieve bitmaps that exactly match
> > >>> existing mappings, where all bits will be written).  Thanks,
> > >>>
> > >>> Alex
> > >>>        
> > >>
> > >> I look at these two ioctls as : VFIO_CHECK_EXTENSION is used to get
> > >> IOMMU type, while VFIO_IOMMU_GET_INFO is used to get properties of a
> > >> particular IOMMU type, right?    
> > > 
> > > Not exclusively, see for example VFIO_DMA_CC_IOMMU,
> > >     
> > >> Then I think VFIO_IOMMU_INFO_DIRTY_PGS should be part of
> > >> VFIO_IOMMU_GET_INFO and when we add code for v2/multi, a flag should be
> > >> added to VFIO_IOMMU_GET_INFO.    
> > > 
> > > Which burns through flags, which is a far more limited resource than
> > > our 32bit extension address space, especially when we're already
> > > planning for one or more extensions to this support.  Thanks,
> > >     
> > 
> > To use flag from VFIO_IOMMU_GET_INFO was your original suggestion, only 
> > 3 bits are used here as of now.  
> 
> Sorry, I'm not infallible.  Perhaps I was short sighted and thought we
> might only need one flag, perhaps I forgot about the check-extension
> ioctl.  Are there any technical reasons to keep it on the get-info
> ioctl?  As I'm trying to look ahead for how we're going to fill the
> gaps of this initial implementation, it seems to me that what we're
> exposing here is in line with what we've used check-extension for in
> the past, and it offers us essentially unlimited extensions to burn
> through, while we're clearly limited on the get-info flags.  We do have
> the precedent of the reset flag on the device_get_info ioctl, but I'm
> largely under the impression that was a mistake and queuing up multiple
> missing features in addition to the base flags feels like compounding
> another mistake.  Thanks,

Another option rather than check-extension that would make sense to me
would be to add a migration capability to the capability chain for the
iommu-get-info ioctl.  We could have our own set of migration related
flags in that capability.  Thanks,

Alex