Re: [kbuild] [linux-next:master 6931/12022] drivers/vfio/vfio_iommu_type1.c:1093 vfio_dma_do_unmap() warn: impossible condition '(size > (~0)) => (0-u32max > u32max)'

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

On Tue, 23 Feb 2021 15:37:31 -0500
Steven Sistare <steven.sistare@oracle.com> wrote:

> On 2/23/2021 12:45 PM, Alex Williamson wrote:
> > On Tue, 23 Feb 2021 08:56:36 -0500
> > Steven Sistare <steven.sistare@oracle.com> wrote:
> >   
> >> On 2/22/2021 6:17 PM, Alex Williamson wrote:  
> >>> On Mon, 22 Feb 2021 15:51:45 -0700
> >>> Alex Williamson <alex.williamson@redhat.com> wrote:
> >>>     
> >>>> On Mon, 22 Feb 2021 17:10:43 +0300
> >>>> Dan Carpenter <dan.carpenter@oracle.com> wrote:
> >>>>    
> >>>>> tree:   https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git  master
> >>>>> head:   37dfbfbdca66834bc0f64ec9b35e09ac6c8898da
> >>>>> commit: 0f53afa12baec8c00f5d1d6afb49325ada105253 [6931/12022] vfio/type1: unmap cleanup      
> >>>>
> >>>> It's always the patches that claim no functional change... ;)
> >>>>    
> >>>>> config: i386-randconfig-m021-20210222 (attached as .config)
> >>>>> compiler: gcc-9 (Debian 9.3.0-15) 9.3.0
> >>>>>
> >>>>> If you fix the issue, kindly add following tag as appropriate
> >>>>> Reported-by: kernel test robot <lkp@intel.com>
> >>>>> Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
> >>>>>
> >>>>> New smatch warnings:
> >>>>> drivers/vfio/vfio_iommu_type1.c:1093 vfio_dma_do_unmap() warn: impossible condition '(size > (~0)) => (0-u32max > u32max)'
> >>>>>
> >>>>> vim +1093 drivers/vfio/vfio_iommu_type1.c
> >>>>>
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1071  static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1072  			     struct vfio_iommu_type1_dma_unmap *unmap,
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1073  			     struct vfio_bitmap *bitmap)
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1074  {
> >>>>> c086de818dd81c Kirti Wankhede  2016-11-17  1075  	struct vfio_dma *dma, *dma_last = NULL;
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1076  	size_t unmapped = 0, pgsize;
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29  1077  	int ret = -EINVAL, retries = 0;
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1078  	unsigned long pgshift;
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29  1079  	dma_addr_t iova = unmap->iova;
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29  1080  	unsigned long size = unmap->size;
> >>>>>                                                         ^^^^^^^^^^^^^^^^^^
> >>>>>
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1081  
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1082  	mutex_lock(&iommu->lock);
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1083  
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1084  	pgshift = __ffs(iommu->pgsize_bitmap);
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1085  	pgsize = (size_t)1 << pgshift;
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1086  
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29  1087  	if (iova & (pgsize - 1))
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1088  		goto unlock;
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1089  
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29  1090  	if (!size || size & (pgsize - 1))
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1091  		goto unlock;
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1092  
> >>>>> 0f53afa12baec8 Steve Sistare   2021-01-29 @1093  	if (iova + size - 1 < iova || size > SIZE_MAX)
> >>>>>
> >>>>> size is unsigned long and SIZE_MAX is ULONG_MAX so "size > SIZE_MAX"
> >>>>> does not make sense.      
> >>>>
> >>>> I think it made sense before the above commit, where unmap->size is a
> >>>> __u64 and a user could provide a value that exceeds SIZE_MAX on ILP32.
> >>>> Seems like the fix is probably to use a size_t for the local variable
> >>>> and restore this test to compare (unmap->size > SIZE_MAX).  Steve?    
> >>>
> >>> Actually it seems like VFIO_DMA_UNMAP_FLAG_ALL doesn't work when
> >>> PHYS_ADDR_MAX != SIZE_MAX (ex. x86 PAE - I think).      
> >>
> >> It seems like PAE causes problems even before VFIO_DMA_UNMAP_FLAG_ALL.  
> > 
> > This wouldn't surprise me, I don't know of any actual non-64bit users
> > and pure 32bit support was only lightly validated ages ago.
> >   
> >> In the previous vfio_dma_do_unmap code, the u64 unmap->size would be
> >> truncated when passed to vfio_find_dma.  
> > 
> > We would have failed with -EINVAL before we get there due to this
> > SIZE_MAX test.  I think the existing (previous) PAE interface is at
> > least self consistent; I see the mapping path also attempts to check
> > that casting map->size as size_t still matches the original value.  
> 
> Good point, and it also checks for vaddr and iova overflow and wrap:
> 
> vfio_dma_do_map()
>         if (map->size != size || map->vaddr != vaddr || map->iova != iova)
>                 return -EINVAL;
>         if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
>                 ret = -EINVAL;
> 
> With that, I don't see a problem with PAE, for unmap-all or otherwise.
> We just need "u64 size" in vfio_dma_do_unmap to avoid the smatch warning.

I'm not convinced.  My understanding is that on PAE phys_addr_t is
64-bit while size_t is 32-bit.  dma_addr_t (iova above) seems to follow
phys_addr_t.  That suggests to me that our {un}map.iova lives in a
64-bit address space, but each mapping is limited to 32-bits.  The
unmap-all logic only looks for a first entry to unmap in the
[0..SIZE_MAX] range.  If an entry happens to exist there, we'll unmap
everything, but the user would have no requirement to have a mapping
within that range, their first mapping could exist at iova = (SIZE_MAX
+ 1).  So unmap-all would effectively need a special case to use
rb_first(), which mostly negates the reason we added
vfio_find_dma_first_node().  Right?  Thanks,

Alex


> >> For unmap, these fixes should suffice, and I would rather do this than
> >> disable the unmap-all flag for a corner case:
> >>
> >>   vfio_dma_do_unmap()
> >>     size_t unmapped = 0;
> >>     unsigned long size = unmap->size;  
> >>     ==>    
> >>     u64 unmapped = 0;
> >>     u64 size = unmap->size;
> >>
> >>   static struct rb_node *vfio_find_dma_first_node(
> >>       struct vfio_iommu *iommu, dma_addr_t start, size_t size)  
> >>   ==>    
> >>   static struct rb_node *vfio_find_dma_first_node(
> >>       struct vfio_iommu *iommu, dma_addr_t start, u64 size)
> >>
> >> And maybe use dma_addr_t instead of u64 in the above (which is 64 bits for
> >> CONFIG_X86_PAE).
> >>
> >> However, there are other places in the existing code that need tweaking
> >> to be safe for PAE, the vfio_find_dma() size arg for one.  
> > 
> > Yes, it looks like the IOMMU aperture checking using vfio_find_dma()
> > could have issues where dma_addr_t > size_t.  Do you want to propose a
> > patch?  Thanks,
> > 
> > Alex
> >   
> >>> I can't say I'm
> >>> really interested in adding complexity to make it work in such a case
> >>> either.  Maybe we can just not expose it, ex:
> >>>
> >>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> >>> index ed03f3fcb07e..6b69a74b3db0 100644
> >>> --- a/drivers/vfio/vfio_iommu_type1.c
> >>> +++ b/drivers/vfio/vfio_iommu_type1.c
> >>> @@ -1207,7 +1207,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>  	int ret = -EINVAL, retries = 0;
> >>>  	unsigned long pgshift;
> >>>  	dma_addr_t iova = unmap->iova;
> >>> -	unsigned long size = unmap->size;
> >>> +	size_t size = unmap->size;
> >>>  	bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
> >>>  	bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
> >>>  	struct rb_node *n, *first_n;
> >>> @@ -1228,7 +1228,7 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
> >>>  		goto unlock;
> >>>  	}
> >>>  
> >>> -	if (iova + size - 1 < iova || size > SIZE_MAX)
> >>> +	if (iova + size - 1 < iova || unmap->size > SIZE_MAX)
> >>>  		goto unlock;
> >>>  
> >>>  	/* When dirty tracking is enabled, allow only min supported pgsize */
> >>> @@ -2657,9 +2657,10 @@ 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_UNMAP_ALL:
> >>>  	case VFIO_UPDATE_VADDR:
> >>>  		return 1;
> >>> +	case VFIO_UNMAP_ALL:
> >>> +		return PHYS_ADDR_MAX == SIZE_MAX ? 1 : 0;
> >>>  	case VFIO_DMA_CC_IOMMU:
> >>>  		if (!iommu)
> >>>  			return 0;
> >>> @@ -2868,6 +2869,10 @@ static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
> >>>  			    VFIO_DMA_UNMAP_FLAG_VADDR)))
> >>>  		return -EINVAL;
> >>>  
> >>> +	if ((PHYS_ADDR_MAX != SIZE_MAX) &&
> >>> +	    (unmap.flags & VFIO_DMA_UNMAP_FLAG_ALL))
> >>> +		return -EINVAL;
> >>> +
> >>>  	if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
> >>>  		unsigned long pgshift;
> >>>  
> >>>
> >>>  
> >>>
> >>>     
> >>>>> Is the " - 1" intentional on the other overflow check?  As in it's okay
> >>>>> to wrap around to zero but not further than that?  Sometimes this is
> >>>>> intentional but it requires more subsystem expertise than I possess.      
> >>>>
> >>>> Yes, since we're dealing with a start + length we need to account for
> >>>> the -1 in the end value, otherwise the user could never unmap the last
> >>>> page of the address space.  Thanks for the report!
> >>>>
> >>>> Alex
> >>>>    
> >>>>> cade075f265b25 Kirti Wankhede  2020-05-29  1094  		goto unlock;
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1095  
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1096  	/* When dirty tracking is enabled, allow only min supported pgsize */
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1097  	if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1098  	    (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1099  		goto unlock;
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1100  	}
> >>>>> 73fa0d10d077d9 Alex Williamson 2012-07-31  1101  
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1102  	WARN_ON((pgsize - 1) & PAGE_MASK);
> >>>>> 331e33d2960c82 Kirti Wankhede  2020-05-29  1103  again:
> >>>>> 1ef3e2bc04223f Alex Williamson 2014-02-26  1104  	/*
> >>>>>
> >>>>> ---
> >>>>> 0-DAY CI Kernel Test Service, Intel Corporation
> >>>>> https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org       
> >>>>    
> >>>     
> >>  
> >   
>