Re: [PATCH V1 4/5] vfio: VA suspend interface

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

On Tue, 12 Jan 2021 15:47:56 -0700
Alex Williamson <alex.williamson@redhat.com> wrote:

> On Mon, 11 Jan 2021 16:15:02 -0500
> Steven Sistare <steven.sistare@oracle.com> wrote:
> 
> > On 1/8/2021 4:15 PM, Alex Williamson wrote:  
> > > On Tue,  5 Jan 2021 07:36:52 -0800
> > > Steve Sistare <steven.sistare@oracle.com> wrote:
> > >     
> > >> Add interfaces that allow the underlying memory object of an iova
> > >> range to be mapped to a new host virtual address in the host process:
> > >>
> > >>   - VFIO_DMA_UNMAP_FLAG_SUSPEND for VFIO_IOMMU_UNMAP_DMA
> > >>   - VFIO_DMA_MAP_FLAG_RESUME flag for VFIO_IOMMU_MAP_DMA
> > >>   - VFIO_SUSPEND extension for VFIO_CHECK_EXTENSION    
> > > 
> > > Suspend and Resume can imply many things other than what's done here.
> > > Should these be something more akin to INVALIDATE_VADDR and
> > > REPLACE_VADDR?    
> > 
> > Agreed.  I suspected we would discuss the names.  Some possibilities:
> > 
> > INVALIDATE_VADDR  REPLACE_VADDR
> > INV_VADDR         SET_VADDR
> > CLEAR_VADDR       SET_VADDR
> > SUSPEND_VADDR     RESUME_VADDR
> >   
> > >> The suspend interface blocks vfio translation of host virtual
> > >> addresses in a range, but DMA to already-mapped pages continues.
> > >> The resume interface records the new base VA and resumes translation.
> > >> See comments in uapi/linux/vfio.h for more details.
> > >>
> > >> This is a partial implementation.  Blocking is added in the next patch.
> > >>
> > >> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
> > >> ---
> > >>  drivers/vfio/vfio_iommu_type1.c | 47 +++++++++++++++++++++++++++++++++++------
> > >>  include/uapi/linux/vfio.h       | 16 ++++++++++++++
> > >>  2 files changed, 57 insertions(+), 6 deletions(-)
> > >>
> > >> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> > >> index 3dc501d..2c164a6 100644
> > >> --- a/drivers/vfio/vfio_iommu_type1.c
> > >> +++ b/drivers/vfio/vfio_iommu_type1.c
> > >> @@ -92,6 +92,7 @@ struct vfio_dma {
> > >>  	int			prot;		/* IOMMU_READ/WRITE */
> > >>  	bool			iommu_mapped;
> > >>  	bool			lock_cap;	/* capable(CAP_IPC_LOCK) */
> > >> +	bool			suspended;    
> > > 
> > > Is there a value we could use for vfio_dma.vaddr that would always be
> > > considered invalid, ex. ULONG_MAX?      
> > 
> > Yes, that could replace the suspend flag.  That, plus changing the language from suspend
> > to invalidate, will probably yield equally understandable code.  I'll try it.  
> 
> Thinking about this further, if we defined a VFIO_IOMMU_TYPE1_INV_VADDR
> as part of the uapi, could we implement this with only a single flag on
> the DMA_MAP ioctl?  For example the user would call DMA_MAP with a flag
> to set the vaddr, first to the invalid valid, then to a new value.  It's
> always seemed a bit awkward to use DMA_UNMAP to invalidate the vaddr
> when the mapping is not actually unmapped.  That might lean towards an
> UPDATE or REPLACE flag.

I realized you really want to make use of the DMA_UNMAP ioctl in order
to use ranges, maybe we can make the mental model more coherent with an
unmap flag like VFIO_DMA_UNMAP_FLAG_VADDR_ONLY, ie. we're only asking
to unmap the vaddr.  The DMA_MAP side might take a similar VADDR_ONLY
flag to reset the vaddr.  That also retains your desired semantics that
we can't "resume" a vaddr that wasn't previously "suspended", we can't
map a vaddr that wasn't previously unmapped.

For the unmap-all problem, userspace already needs to work around this,
see for instance QEMU:

1b296c3def4b vfio: Don't issue full 2^64 unmap

So I wonder really how critical it is and whether it really would be
sufficient for userspace to track a high water mark for mappings.
Otherwise, I think I'm leaning towards a DMA_UNMAP flag like
VFIO_DMA_UNMAP_FLAG_UNMAP_ALL that would disregard the iova and size
fields to apply to all mappings.  Designating special values for iova or
size trigger extended behavior feels a bit hackish.  Thanks,

Alex
 
> > > We'd need to decide if we want to
> > > allow users to create mappings (mdev-only) using an initial invalid
> > > vaddr.    
> > 
> > Maybe.  Not sure yet.  
> 
> If we used the above, it almost seems strange not to allow it, but at
> the same time we don't really want to have different rules for
> different devices types.  An initially valid vaddr doesn't seem
> unreasonable... though we don't test it until the vendor driver tries
> to pin or rw pages w/o IOMMU backing.
>  
> > >>  	struct task_struct	*task;
> > >>  	struct rb_root		pfn_list;	/* Ex-user pinned pfn list */
> > >>  	unsigned long		*bitmap;
> > >> @@ -1080,7 +1081,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> > >>  	int ret = 0, retries = 0;
> > >>  	unsigned long pgshift;
> > >>  	dma_addr_t iova;
> > >> -	unsigned long size;
> > >> +	unsigned long size, consumed;    
> > > 
> > > This could be scoped into the branch below.    
> > 
> > OK.
> >   
> > >>  	mutex_lock(&iommu->lock);
> > >>  
> > >> @@ -1169,6 +1170,21 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> > >>  		if (dma->task->mm != current->mm)
> > >>  			break;
> > >>  
> > >> +		if (unmap->flags & VFIO_DMA_UNMAP_FLAG_SUSPEND) {
> > >> +			if (dma->suspended) {
> > >> +				ret = -EINVAL;
> > >> +				goto unlock;
> > >> +			}    
> > > 
> > > This leaves us in a state where we marked some entries but not others.
> > > We should either unwind or... what's the actual harm in skipping these?    
> > 
> > We could skip them with no ill effect.  However, it likely means the app is confused
> > and potentially broken, and it would be courteous to inform them so.  I found such bugs
> > in qemu as I was developing this feature.
> > 
> > IMO unwinding does not help the app, and adds unnecessary code.  It can still leave some
> > ranges suspended and some not.  The safest recovery is for the app to exit, and tell the 
> > developer to fix the redundant suspend call.  
> 
> That sounds like an entirely practical rationalization, but our
> standard practice is to maintain a consistent state.  If an ioctl fails
> is should effectively be as if the ioctl was never called, where
> possible.  Userspace can be broken, and potentially so broken that their
> best choice is to abort, but we should maintain consistent, predictable
> behavior.
> 
> > >> +			dma->suspended = true;
> > >> +			consumed = dma->iova + dma->size - iova;
> > >> +			if (consumed >= size)
> > >> +				break;
> > >> +			iova += consumed;
> > >> +			size -= consumed;
> > >> +			unmapped += dma->size;
> > >> +			continue;
> > >> +		}    
> > > 
> > > This short-cuts the dirty bitmap flag, so we need to decide if it's
> > > legal to call them together or we need to prevent it... Oh, I see
> > > you've excluded them earlier below.
> > >     
> > >> +
> > >>  		if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
> > >>  			struct vfio_iommu_type1_dma_unmap nb_unmap;
> > >>  
> > >> @@ -1307,6 +1323,7 @@ static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
> > >>  static int vfio_dma_do_map(struct vfio_iommu *iommu,
> > >>  			   struct vfio_iommu_type1_dma_map *map)
> > >>  {
> > >> +	bool resume = map->flags & VFIO_DMA_MAP_FLAG_RESUME;
> > >>  	dma_addr_t iova = map->iova;
> > >>  	unsigned long vaddr = map->vaddr;
> > >>  	size_t size = map->size;
> > >> @@ -1324,13 +1341,16 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
> > >>  	if (map->flags & VFIO_DMA_MAP_FLAG_READ)
> > >>  		prot |= IOMMU_READ;
> > >>  
> > >> +	if ((prot && resume) || (!prot && !resume))
> > >> +		return -EINVAL;
> > >> +
> > >>  	mutex_lock(&iommu->lock);
> > >>  
> > >>  	pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
> > >>  
> > >>  	WARN_ON((pgsize - 1) & PAGE_MASK);
> > >>  
> > >> -	if (!prot || !size || (size | iova | vaddr) & (pgsize - 1)) {
> > >> +	if (!size || (size | iova | vaddr) & (pgsize - 1)) {
> > >>  		ret = -EINVAL;
> > >>  		goto out_unlock;
> > >>  	}
> > >> @@ -1341,7 +1361,19 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
> > >>  		goto out_unlock;
> > >>  	}
> > >>  
> > >> -	if (vfio_find_dma(iommu, iova, size)) {
> > >> +	dma = vfio_find_dma(iommu, iova, size);
> > >> +	if (resume) {
> > >> +		if (!dma) {
> > >> +			ret = -ENOENT;
> > >> +		} else if (!dma->suspended || dma->iova != iova ||
> > >> +			   dma->size != size) {    
> > > 
> > > Why is it necessary that the vfio_dma be suspended before being
> > > resumed?  Couldn't a user simply use this to change the vaddr?  Does
> > > that promote abusive use?    
> > 
> > This would almost always be incorrect.  If the vaddr changes, then the old vaddr was already
> > invalidated, and there is a window where it is not OK for kernel code to use the old vaddr.
> > This could only be safe if the memory object is mapped at both the old vaddr and the new
> > vaddr concurrently, which is an unlikely use case.  
> 
> Ok, it's not like the use can't make it instantaneously invalid and then
> replace it.
> 
> > >> +			ret = -EINVAL;
> > >> +		} else {
> > >> +			dma->vaddr = vaddr;    
> > > 
> > > Seems like there's a huge opportunity for a user to create coherency
> > > issues here... it's their data though I guess.    
> > 
> > Yes.  That's what the language in the uapi about mapping the same memory object is about.
> >   
> > >> +			dma->suspended = false;
> > >> +		}
> > >> +		goto out_unlock;
> > >> +	} else if (dma) {
> > >>  		ret = -EEXIST;
> > >>  		goto out_unlock;
> > >>  	}
> > >> @@ -2532,6 +2564,7 @@ static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
> > >>  	case VFIO_TYPE1_IOMMU:
> > >>  	case VFIO_TYPE1v2_IOMMU:
> > >>  	case VFIO_TYPE1_NESTING_IOMMU:
> > >> +	case VFIO_SUSPEND:
> > >>  		return 1;
> > >>  	case VFIO_DMA_CC_IOMMU:
> > >>  		if (!iommu)
> > >> @@ -2686,7 +2719,8 @@ static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
> > >>  {
> > >>  	struct vfio_iommu_type1_dma_map map;
> > >>  	unsigned long minsz;
> > >> -	uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
> > >> +	uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
> > >> +			VFIO_DMA_MAP_FLAG_RESUME;
> > >>  
> > >>  	minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
> > >>  
> > >> @@ -2704,6 +2738,8 @@ static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
> > >>  {
> > >>  	struct vfio_iommu_type1_dma_unmap unmap;
> > >>  	struct vfio_bitmap bitmap = { 0 };
> > >> +	uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
> > >> +			VFIO_DMA_UNMAP_FLAG_SUSPEND;
> > >>  	unsigned long minsz;
> > >>  	int ret;
> > >>  
> > >> @@ -2712,8 +2748,7 @@ static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
> > >>  	if (copy_from_user(&unmap, (void __user *)arg, minsz))
> > >>  		return -EFAULT;
> > >>  
> > >> -	if (unmap.argsz < minsz ||
> > >> -	    unmap.flags & ~VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP)
> > >> +	if (unmap.argsz < minsz || unmap.flags & ~mask || unmap.flags == mask)    
> > > 
> > > Maybe a short comment here to note that dirty-bimap and
> > > suspend/invalidate are mutually exclusive.  Probably should be
> > > mentioned in the uapi too.    
> > 
> > Will do, for both.
> >   
> > >>  		return -EINVAL;
> > >>  
> > >>  	if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
> > >> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
> > >> index 896e527..fcf7b56 100644
> > >> --- a/include/uapi/linux/vfio.h
> > >> +++ b/include/uapi/linux/vfio.h
> > >> @@ -46,6 +46,9 @@
> > >>   */
> > >>  #define VFIO_NOIOMMU_IOMMU		8
> > >>  
> > >> +/* Supports VFIO DMA suspend and resume */
> > >> +#define VFIO_SUSPEND			9
> > >> +
> > >>  /*
> > >>   * The IOCTL interface is designed for extensibility by embedding the
> > >>   * structure length (argsz) and flags into structures passed between
> > >> @@ -1046,12 +1049,19 @@ struct vfio_iommu_type1_info_cap_migration {
> > >>   *
> > >>   * Map process virtual addresses to IO virtual addresses using the
> > >>   * provided struct vfio_dma_map. Caller sets argsz. READ &/ WRITE required.
> > >> + *
> > >> + * If flags & VFIO_DMA_MAP_FLAG_RESUME, record the new base vaddr for iova, and
> > >> + * resume translation of host virtual addresses in the iova range.  The new
> > >> + * vaddr must point to the same memory object as the old vaddr, but this is not
> > >> + * verified.    
> > > 
> > > It's hard to use "must" terminology here if we're not going to check.
> > > Maybe the phrasing should be something more along the lines of "should
> > > point to the same memory object or the user risks coherency issues
> > > within their virtual address space".    
> > 
> > I used "must" because it is always incorrect if the object is not the same.  How about:
> >   The new vaddr must point to the same memory object as the old vaddr, but this is not
> >   verified.  Violation of this constraint may result in memory corruption within the
> >   host process and/or guest.  
> 
> Since the "must" is not relative to the API but to the resulting
> behavior, perhaps something like:
> 
>   In order to maintain memory consistency within the user application,
>   the updated vaddr must address the same memory object as originally
>   mapped, failure to do so will result in user memory corruption and/or
>   device misbehavior.
> 
> Thanks,
> Alex
> 
> > >>  iova and size must match those in the original MAP_DMA call.
> > >> + * Protection is not changed, and the READ & WRITE flags must be 0.    
> > > 
> > > This doesn't mention that the entry must be previously
> > > suspended/invalidated (if we choose to keep those semantics).  Thanks,    
> > 
> > Will add, thanks.
> > 
> > - Steve   
> > >>   */
> > >>  struct vfio_iommu_type1_dma_map {
> > >>  	__u32	argsz;
> > >>  	__u32	flags;
> > >>  #define VFIO_DMA_MAP_FLAG_READ (1 << 0)		/* readable from device */
> > >>  #define VFIO_DMA_MAP_FLAG_WRITE (1 << 1)	/* writable from device */
> > >> +#define VFIO_DMA_MAP_FLAG_RESUME (1 << 2)
> > >>  	__u64	vaddr;				/* Process virtual address */
> > >>  	__u64	iova;				/* IO virtual address */
> > >>  	__u64	size;				/* Size of mapping (bytes) */
> > >> @@ -1084,11 +1094,17 @@ struct vfio_bitmap {
> > >>   * indicates that the page at that offset from iova is dirty. A Bitmap of the
> > >>   * pages in the range of unmapped size is returned in the user-provided
> > >>   * vfio_bitmap.data.
> > >> + *
> > >> + * If flags & VFIO_DMA_UNMAP_FLAG_SUSPEND, do not unmap, but suspend vfio
> > >> + * translation of host virtual addresses in the iova range.  During suspension,
> > >> + * kernel threads that attempt to translate will block.  DMA to already-mapped
> > >> + * pages continues.
> > >>   */
> > >>  struct vfio_iommu_type1_dma_unmap {
> > >>  	__u32	argsz;
> > >>  	__u32	flags;
> > >>  #define VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP (1 << 0)
> > >> +#define VFIO_DMA_UNMAP_FLAG_SUSPEND	     (1 << 1)
> > >>  	__u64	iova;				/* IO virtual address */
> > >>  	__u64	size;				/* Size of mapping (bytes) */
> > >>  	__u8    data[];    
> > >     
> >   
>