Re: [PATCH V1 2/5] vfio: option to unmap all

From: Steven Sistare <steven.sistare@oracle.com>
To: Alex Williamson <alex.williamson@redhat.com>
Cc: kvm@vger.kernel.org, Cornelia Huck <cohuck@redhat.com>,
	Kirti Wankhede <kwankhede@nvidia.com>
Subject: Re: [PATCH V1 2/5] vfio: option to unmap all
Date: Mon, 11 Jan 2021 16:09:18 -0500	[thread overview]
Message-ID: <c5850c5d-52c4-1f21-30cc-34f9b2d7b7f3@oracle.com> (raw)
In-Reply-To: <20210108123548.033377e7@omen.home>

On 1/8/2021 2:35 PM, Alex Williamson wrote:
> Hi Steve,
> 
> On Tue,  5 Jan 2021 07:36:50 -0800
> Steve Sistare <steven.sistare@oracle.com> wrote:
> 
>> For VFIO_IOMMU_UNMAP_DMA, delete all mappings if iova=0 and size=0.
> 
> Only the latter is invalid, iova=0 is not special, so does it make
> sense to use this combination to invoke something special?  It seems
> like it opens the door that any size less than the minimum mapping
> granularity means something special.
> 
> Why not use a flag to trigger an unmap-all?

Hi Alex, that would be fine.

> Does userspace have any means to know this is supported other than to
> test it before creating any mappings?

Not currently.  We could overload VFIO_SUSPEND, or define a new extension code.

> What's the intended interaction with retrieving the dirty bitmap during
> an unmap-all?

Undefined and broken if there are gaps between segments :(  Good catch, thanks.  
I will disallow the combination of unmap-all and get-dirty-bitmap.

>> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
>> ---
>>  drivers/vfio/vfio_iommu_type1.c | 11 ++++++++---
>>  include/uapi/linux/vfio.h       |  3 ++-
>>  2 files changed, 10 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>> index 02228d0..3dc501d 100644
>> --- a/drivers/vfio/vfio_iommu_type1.c
>> +++ b/drivers/vfio/vfio_iommu_type1.c
>> @@ -1079,6 +1079,8 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
>>  	size_t unmapped = 0, pgsize;
>>  	int ret = 0, retries = 0;
>>  	unsigned long pgshift;
>> +	dma_addr_t iova;
>> +	unsigned long size;
>>  
>>  	mutex_lock(&iommu->lock);
>>  
>> @@ -1090,7 +1092,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
>>  		goto unlock;
>>  	}
>>  
>> -	if (!unmap->size || unmap->size & (pgsize - 1)) {
>> +	if ((!unmap->size && unmap->iova) || unmap->size & (pgsize - 1)) {
>>  		ret = -EINVAL;
>>  		goto unlock;
>>  	}
>> @@ -1154,8 +1156,11 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> 
> It looks like the code just above this would have an issue if there are
> dma mappings at iova=0.

Are you referring to this code?

        if (iommu->v2) {
                dma = vfio_find_dma(iommu, unmap->iova, 1);
                if (dma && dma->iova != unmap->iova) {
                        ret = -EINVAL;

Both unmap->iova and dma->iova would be 0, so I don't see the problem.

>>  		}
>>  	}
>>  
>> -	while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
>> -		if (!iommu->v2 && unmap->iova > dma->iova)
>> +	iova = unmap->iova;
>> +	size = unmap->size ? unmap->size : SIZE_MAX;
> 
> AFAICT the only difference of this versus the user calling the unmap
> with iova=0 size=SIZE_MAX is that SIZE_MAX will throw an -EINVAL due to
> page size alignment.  If we assume there are no IOMMUs with 1 byte page
> size, the special combination could instead be {0, SIZE_MAX}.  

Fine, but we would still need to document it specifically so the user knows that 
the unaligned SIZE_MAX does not return EINVAL.

> Or the
> caller could just track a high water mark for their mappings and use
> the interface that exists.  Thanks,

I am trying to avoid the need to modify existing code, for legacy qemu live update.
Either a new flag or {0, SIZE_MAX} is suitable.  Which do you prefer?

- Steve

>> +
>> +	while ((dma = vfio_find_dma(iommu, iova, size))) {
>> +		if (!iommu->v2 && iova > dma->iova)
>>  			break;
>>  		/*
>>  		 * Task with same address space who mapped this iova range is
>> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
>> index 9204705..896e527 100644
>> --- a/include/uapi/linux/vfio.h
>> +++ b/include/uapi/linux/vfio.h
>> @@ -1073,7 +1073,8 @@ struct vfio_bitmap {
>>   * Caller sets argsz.  The actual unmapped size is returned in the size
>>   * field.  No guarantee is made to the user that arbitrary unmaps of iova
>>   * or size different from those used in the original mapping call will
>> - * succeed.
>> + * succeed.  If iova=0 and size=0, all addresses are unmapped.
>> + *
>>   * VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP should be set to get the dirty bitmap
>>   * before unmapping IO virtual addresses. When this flag is set, the user must
>>   * provide a struct vfio_bitmap in data[]. User must provide zero-allocated
> 