Re: [PATCH v16 Kernel 5/7] vfio iommu: Update UNMAP_DMA ioctl to get dirty bitmap before unmap

[Yan Zhao]

On Fri, Mar 27, 2020 at 12:42:43PM +0800, Kirti Wankhede wrote:
> 
> 
> On 3/27/2020 5:34 AM, Yan Zhao wrote:
> > On Fri, Mar 27, 2020 at 05:39:44AM +0800, Kirti Wankhede wrote:
> >>
> >>
> >> On 3/25/2020 7:48 AM, Yan Zhao wrote:
> >>> On Wed, Mar 25, 2020 at 03:32:37AM +0800, Kirti Wankhede wrote:
> >>>> DMA mapped pages, including those pinned by mdev vendor drivers, might
> >>>> get unpinned and unmapped while migration is active and device is still
> >>>> running. For example, in pre-copy phase while guest driver could access
> >>>> those pages, host device or vendor driver can dirty these mapped pages.
> >>>> Such pages should be marked dirty so as to maintain memory consistency
> >>>> for a user making use of dirty page tracking.
> >>>>
> >>>> To get bitmap during unmap, user should allocate memory for bitmap, set
> >>>> size of allocated memory, set page size to be considered for bitmap and
> >>>> set flag VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP.
> >>>>
> >>>> Signed-off-by: Kirti Wankhede <address@hidden>
> >>>> Reviewed-by: Neo Jia <address@hidden>
> >>>> ---
> >>>>    drivers/vfio/vfio_iommu_type1.c | 54 
> >>>> ++++++++++++++++++++++++++++++++++++++---
> >>>>    include/uapi/linux/vfio.h       | 10 ++++++++
> >>>>    2 files changed, 60 insertions(+), 4 deletions(-)
> >>>>
> >>>> diff --git a/drivers/vfio/vfio_iommu_type1.c 
> >>>> b/drivers/vfio/vfio_iommu_type1.c
> >>>> index 27ed069c5053..b98a8d79e13a 100644
> >>>> --- a/drivers/vfio/vfio_iommu_type1.c
> >>>> +++ b/drivers/vfio/vfio_iommu_type1.c
> >>>> @@ -982,7 +982,8 @@ static int verify_bitmap_size(uint64_t npages, 
> >>>> uint64_t bitmap_size)
> >>>>    }
> >>>>    
> >>>>    static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>> -                             struct vfio_iommu_type1_dma_unmap *unmap)
> >>>> +                             struct vfio_iommu_type1_dma_unmap *unmap,
> >>>> +                             struct vfio_bitmap *bitmap)
> >>>>    {
> >>>>          uint64_t mask;
> >>>>          struct vfio_dma *dma, *dma_last = NULL;
> >>>> @@ -1033,6 +1034,10 @@ static int vfio_dma_do_unmap(struct vfio_iommu 
> >>>> *iommu,
> >>>>           * will be returned if these conditions are not met.  The v2 
> >>>> interface
> >>>>           * will only return success and a size of zero if there were no
> >>>>           * mappings within the range.
> >>>> +         *
> >>>> +         * When VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP flag is set, unmap 
> >>>> request
> >>>> +         * must be for single mapping. Multiple mappings with this flag 
> >>>> set is
> >>>> +         * not supported.
> >>>>           */
> >>>>          if (iommu->v2) {
> >>>>                  dma = vfio_find_dma(iommu, unmap->iova, 1);
> >>>> @@ -1040,6 +1045,13 @@ static int vfio_dma_do_unmap(struct vfio_iommu 
> >>>> *iommu,
> >>>>                          ret = -EINVAL;
> >>>>                          goto unlock;
> >>>>                  }
> >>>> +
> >>>> +                if ((unmap->flags & 
> >>>> VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
> >>>> +                    (dma->iova != unmap->iova || dma->size != 
> >>>> unmap->size)) {
> >>> potential NULL pointer!
> >>>
> >>> And could you address the comments in v14?
> >>> How to handle DSI unmaps in vIOMMU
> >>> (https://lore.kernel.org/kvm/20200323011041.GB5456@joy-OptiPlex-7040/)
> >>>
> >>
> >> Sorry, I drafted reply to it, but I missed to send, it remained in my 
> >> drafts
> >>
> >>   >
> >>   > it happens in vIOMMU Domain level invalidation of IOTLB
> >>   > (domain-selective invalidation, see vtd_iotlb_domain_invalidate() in
> >> qemu).
> >>   > common in VTD lazy mode, and NOT just happening once at boot time.
> >>   > rather than invalidate page by page, it batches the page invalidation.
> >>   > so, when this invalidation takes place, even higher level page tables
> >>   > have been invalid and therefore it has to invalidate a bigger
> >> combined range.
> >>   > That's why we see IOVAs are mapped in 4k pages, but are unmapped in 2M
> >>   > pages.
> >>   >
> >>   > I think those UNMAPs should also have GET_DIRTY_BIMTAP flag on, right?
> >>
> >>
> >> vtd_iotlb_domain_invalidate()
> >>     vtd_sync_shadow_page_table()
> >>       vtd_sync_shadow_page_table_range(vtd_as, &ce, 0, UINT64_MAX)
> >>         vtd_page_walk()
> >>           vtd_page_walk_level() - walk over specific level for IOVA range
> >>             vtd_page_walk_one()
> >>               memory_region_notify_iommu()
> >>               ...
> >>                 vfio_iommu_map_notify()
> >>
> >> In the above trace, isn't page walk will take care of creating proper
> >> IOTLB entry which should be same as created during mapping for that
> >> IOTLB entry?
> >>
> > No. It does walk the page table, but as it's dsi (delay & batched unmap),
> > pages table entry for a whole 2M (the higher level, not last level for 4K)
> > range is invalid, so the iotlb->addr_mask what vfio_iommu_map_notify()
> > receives is (2M - 1), not the same as the size for map.
> > 
> 
> When do this happen? during my testing I never hit this case. How can I 
> hit this case?

Just common settings to enable vIOMMU:
Qemu: -device intel-iommu,caching-mode=true
guest kernel parameter: intel_iommu=on

(intel_iommu=on turns on lazy mode by default)

In lazy mode, guest notifies DMA MAP on page level, but notifies DMA UNMAPs
in batch.
with a pass-through NVMe, there are 89 DSI unmaps in 1 second for a typical fio.
With a pass-through GPU, there 22 DSI unmaps in total for benchmark openArena
(lasting around 55 secs)
> 
> In this case, will adjacent whole vfio_dmas will be clubbed together or 
> will there be any intersection of vfio_dmas?
>
clubbed together.