Hi Alex,
On 2020/12/15 8:04, Alex Williamson wrote:
> On Thu, 10 Dec 2020 15:34:22 +0800
> Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>
>> When we pin or detach a group which is not dirty tracking capable,
>> we will try to promote pinned_scope of vfio_iommu.
>>
>> If we succeed to do so, vfio only report pinned_scope as dirty to
>> userspace next time, but these memory written before pin or detach
>> is missed.
>>
>> The solution is that we must populate all dma range as dirty before
>> promoting pinned_scope of vfio_iommu.
>
> Please don't bury fixes patches into a series with other optimizations
> and semantic changes. Send it separately.
>
OK, I will.
>>
>> Signed-off-by: Keqian Zhu <zhukeqian1@xxxxxxxxxx>
>> ---
>> drivers/vfio/vfio_iommu_type1.c | 18 ++++++++++++++++++
>> 1 file changed, 18 insertions(+)
>>
>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>> index bd9a94590ebc..00684597b098 100644
>> --- a/drivers/vfio/vfio_iommu_type1.c
>> +++ b/drivers/vfio/vfio_iommu_type1.c
>> @@ -1633,6 +1633,20 @@ static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
>> return group;
>> }
>>
>> +static void vfio_populate_bitmap_all(struct vfio_iommu *iommu)
>> +{
>> + struct rb_node *n;
>> + unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
>> +
>> + for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
>> + struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
>> + unsigned long nbits = dma->size >> pgshift;
>> +
>> + if (dma->iommu_mapped)
>> + bitmap_set(dma->bitmap, 0, nbits);
>> + }
>> +}
>
>
> If we detach a group which results in only non-IOMMU backed mdevs,
> don't we also clear dma->iommu_mapped as part of vfio_unmap_unpin()
> such that this test is invalid? Thanks,
Good spot :-). The code will skip bitmap_set under this situation.
We should set the bitmap unconditionally when vfio_iommu is promoted,
as we must have IOMMU backed domain before promoting the vfio_iommu.
Besides, I think we should also mark dirty in vfio_remove_dma if dirty
tracking is active. Right?
Thanks,
Keqian
>
> Alex
>
>> +
>> static void promote_pinned_page_dirty_scope(struct vfio_iommu *iommu)
>> {
>> struct vfio_domain *domain;
>> @@ -1657,6 +1671,10 @@ static void promote_pinned_page_dirty_scope(struct vfio_iommu *iommu)
>> }
>>
>> iommu->pinned_page_dirty_scope = true;
>> +
>> + /* Set all bitmap to avoid missing dirty page */
>> + if (iommu->dirty_page_tracking)
>> + vfio_populate_bitmap_all(iommu);
>> }
>>
>> static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
>
> .
>
