Re: [PATCH Kernel v20 6/8] vfio iommu: Update UNMAP_DMA ioctl to get dirty bitmap before unmap

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

On Fri, 15 May 2020 21:00:32 +0530
Kirti Wankhede <kwankhede@nvidia.com> wrote:

> On 5/15/2020 7:01 PM, Alex Williamson wrote:
> > On Fri, 15 May 2020 12:17:03 +0530
> > Kirti Wankhede <kwankhede@nvidia.com> wrote:
> >   
> >> On 5/15/2020 11:17 AM, Alex Williamson wrote:  
> >>> On Fri, 15 May 2020 09:46:43 +0530
> >>> Kirti Wankhede <kwankhede@nvidia.com> wrote:
> >>>      
> >>>> On 5/15/2020 8:57 AM, Alex Williamson wrote:  
> >>>>> On Fri, 15 May 2020 02:07:45 +0530
> >>>>> Kirti Wankhede <kwankhede@nvidia.com> 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
> >>>>>> it all zeros, 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 <kwankhede@nvidia.com>
> >>>>>> Reviewed-by: Neo Jia <cjia@nvidia.com>
> >>>>>> ---
> >>>>>>     drivers/vfio/vfio_iommu_type1.c | 77 ++++++++++++++++++++++++++++++++++-------
> >>>>>>     include/uapi/linux/vfio.h       | 10 ++++++
> >>>>>>     2 files changed, 75 insertions(+), 12 deletions(-)
> >>>>>>
> >>>>>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> >>>>>> index b76d3b14abfd..a1dc57bcece5 100644
> >>>>>> --- a/drivers/vfio/vfio_iommu_type1.c
> >>>>>> +++ b/drivers/vfio/vfio_iommu_type1.c
> >>>>>> @@ -195,11 +195,15 @@ static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
> >>>>>>     static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
> >>>>>>     {
> >>>>>>     	uint64_t npages = dma->size / pgsize;
> >>>>>> +	size_t bitmap_size;
> >>>>>>     
> >>>>>>     	if (npages > DIRTY_BITMAP_PAGES_MAX)
> >>>>>>     		return -EINVAL;
> >>>>>>     
> >>>>>> -	dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages), GFP_KERNEL);
> >>>>>> +	/* Allocate extra 64 bits which are used for bitmap manipulation */
> >>>>>> +	bitmap_size = DIRTY_BITMAP_BYTES(npages) + sizeof(u64);
> >>>>>> +
> >>>>>> +	dma->bitmap = kvzalloc(bitmap_size, GFP_KERNEL);
> >>>>>>     	if (!dma->bitmap)
> >>>>>>     		return -ENOMEM;
> >>>>>>     
> >>>>>> @@ -999,23 +1003,25 @@ 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;
> >>>>>> -	size_t unmapped = 0;
> >>>>>> +	size_t unmapped = 0, pgsize;
> >>>>>>     	int ret = 0, retries = 0;
> >>>>>> +	unsigned long pgshift;
> >>>>>>     
> >>>>>>     	mutex_lock(&iommu->lock);
> >>>>>>     
> >>>>>> -	mask = ((uint64_t)1 << __ffs(iommu->pgsize_bitmap)) - 1;
> >>>>>> +	pgshift = __ffs(iommu->pgsize_bitmap);
> >>>>>> +	pgsize = (size_t)1 << pgshift;
> >>>>>>     
> >>>>>> -	if (unmap->iova & mask) {
> >>>>>> +	if (unmap->iova & (pgsize - 1)) {
> >>>>>>     		ret = -EINVAL;
> >>>>>>     		goto unlock;
> >>>>>>     	}
> >>>>>>     
> >>>>>> -	if (!unmap->size || unmap->size & mask) {
> >>>>>> +	if (!unmap->size || unmap->size & (pgsize - 1)) {
> >>>>>>     		ret = -EINVAL;
> >>>>>>     		goto unlock;
> >>>>>>     	}
> >>>>>> @@ -1026,9 +1032,15 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>>>>     		goto unlock;
> >>>>>>     	}
> >>>>>>     
> >>>>>> -	WARN_ON(mask & PAGE_MASK);
> >>>>>> -again:
> >>>>>> +	/* When dirty tracking is enabled, allow only min supported pgsize */
> >>>>>> +	if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
> >>>>>> +	    (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
> >>>>>> +		ret = -EINVAL;
> >>>>>> +		goto unlock;
> >>>>>> +	}
> >>>>>>     
> >>>>>> +	WARN_ON((pgsize - 1) & PAGE_MASK);
> >>>>>> +again:
> >>>>>>     	/*
> >>>>>>     	 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
> >>>>>>     	 * avoid tracking individual mappings.  This means that the granularity
> >>>>>> @@ -1066,6 +1078,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>>>>     			ret = -EINVAL;
> >>>>>>     			goto unlock;
> >>>>>>     		}
> >>>>>> +
> >>>>>>     		dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
> >>>>>>     		if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
> >>>>>>     			ret = -EINVAL;
> >>>>>> @@ -1083,6 +1096,23 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>>>>     		if (dma->task->mm != current->mm)
> >>>>>>     			break;
> >>>>>>     
> >>>>>> +		if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
> >>>>>> +		    (dma_last != dma)) {
> >>>>>> +
> >>>>>> +			/*
> >>>>>> +			 * mark all pages dirty if all pages are pinned and
> >>>>>> +			 * mapped
> >>>>>> +			 */
> >>>>>> +			if (dma->iommu_mapped)
> >>>>>> +				bitmap_set(dma->bitmap, 0,
> >>>>>> +					   dma->size >> pgshift);  
> >>>>>
> >>>>> Nit, all the callers of update_user_bitmap() precede the call with this
> >>>>> identical operation, we should probably push it into the function to do
> >>>>> it.
> >>>>>         
> >>>>>> +
> >>>>>> +			ret = update_user_bitmap(bitmap->data, dma,
> >>>>>> +						 unmap->iova, pgsize);
> >>>>>> +			if (ret)
> >>>>>> +				break;
> >>>>>> +		}
> >>>>>> +  
> >>>>>
> >>>>> As noted last time, the above is just busy work if pfn_list is not
> >>>>> already empty.  The entire code block above should be moved to after
> >>>>> the block below.  Thanks,
> >>>>>         
> >>>>
> >>>> pfn_list will be empty for IOMMU backed devices where all pages are
> >>>> pinned and mapped,  
> >>>
> >>> Unless we're making use of the selective dirtying introduced in patch
> >>> 8/8 or the container is shared with non-IOMMU backed mdevs.
> >>>      
> >>>> but those should be reported as dirty.  
> >>>
> >>> I'm confused how that justifies or requires this ordering.
> >>>      
> >>
> >> 1. non IOMMU mdev device:
> >> - vendor driver pins pages
> >> - pfn_list is not empty
> >> - device dma or write to pinned pages
> >>
> >> 2. IOMMU backed mdev device or vfio device, but smart driver which pins
> >> required pages
> >> - vendor driver pins pages
> >> - pfn_list is not empty
> >> - device dma or write to pinned pages
> >>
> >> 3. IOMMU backed mdev device or vfio device, driver is not smart
> >> - pages are pinned and mapped during attach
> >> - pfn_list is empty
> >> - device dma or write to any of pinned pages
> >>
> >> For case 3, here this function does bitmap_set(dma->bitmap), that is
> >> mark all pages dirty and then accordingly copy bitmap to user buffer.
> >> Copying dma->bitmap logic remains same.  
> > 
> > What dependency does case 3 have on pfn_list?  Why does it matter if
> > that bitmap_set() occurs before or after we've drained pfn_list?  How
> > do we know we're in case 3 before we've looked at pfn_list?
> >   
> 
> If update_user_bitmap() is moved in if block like:
> if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
> 	update_user_bitmap()
> ...
> }
> 
> then update_user_bitmap() would never get called for case 3.
> 
> >>>> So moved it
> >>>> back above empty pfn_list check.  
> >>>
> >>> Sorry, it still doesn't make any sense to me, and with no discussion I
> >>> can't differentiate ignored comments from discarded comments.
> >>>
> >>> Pages in the pfn_list contribute to the dirty bitmap when they're
> >>> pinned, we don't depend on pfn_list when reporting the dirty bitmap
> >>> except for re-populating pfn_list dirtied pages after the bitmap has
> >>> been cleared.  We're unmapping the dma, so that's not the case here.
> >>> Also since update_user_bitmap() shifts the bitmap in place now, any
> >>> repetitive calls will give us incorrect results.  
> >>
> >> Right, but this is unmapping and freeing vfio_dma
> >>  
> >>>   Therefore, as I see
> >>> it, we _can_ take the branch below and when we do any work we've done
> >>> above is not only wasted but may lead to incorrect data copied to
> >>> the user if we shift dma->bitmap in place more than once.  Please
> >>> explain in more detail if you believe this is still correct.  Thanks,
> >>>      
> >>
> >> In this case also bitmap copy to user happens once, (dma_last != dma)
> >> takes care of making sure that its called only once.  
> > 
> > I did miss the dma_last check, so that prevents us from repeating this
> > path, BUT we release iommu->lock if we enter the pfn_list !empty path.
> > The moment we do that, we might unblock a vendor driver trying to pin
> > or rw more pages.  So the bitmap we've copied to the user might be stale
> > and incorrect.  I don't see why this is so contentious, what is the
> > actual disadvantage to moving this until after we've drained pfn_list?
> > The risks seem abundant to me.  Thanks,
> >   
> 
> Sorry if I misinterpreted earlier, are you proposing to move 
> update_user_bitmap() after this if block?
> 
> if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
> ...
> }
> 
> I thought you are asking to move it in block.
> That makes sense. I'll update the patch in send next version in some time.

Yes, after not in.  Thanks,

Alex