Despite carefully rewording the kerneldoc to describe the new direct
interaction with dma_range_map, it seems I managed to confuse myself in
removing the redundant force_aperture check and ended up making the code
not do that at all. This led to dma_range_maps inadvertently being able
to set iovad->start_pfn = 0, and all the nonsensical chaos which ensues
from there. Restore the correct behaviour of constraining base_pfn to
the domain aperture regardless of dma_range_map, and not trying to apply
dma_range_map constraints to the basic IOVA domain since they will be
properly handled with reserved regions later.

Reported-by: Jon Hunter <jonathanh@nvidia.com>
Reported-by: Jerry Snitselaar <jsnitsel@redhat.com>
Fixes: e28a114243d1 ("iommu/dma: Make limit checks self-contained")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---
 drivers/iommu/dma-iommu.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 89a53c2f2cf9..1ec89ca75c59 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -685,15 +685,15 @@ static int iommu_dma_init_domain(struct iommu_domain *domain, struct device *dev
 
 	/* Check the domain allows at least some access to the device... */
 	if (map) {
-		dma_addr_t base = dma_range_map_min(map);
-		if (base > domain->geometry.aperture_end ||
+		if (dma_range_map_min(map) > domain->geometry.aperture_end ||
 		    dma_range_map_max(map) < domain->geometry.aperture_start) {
 			pr_warn("specified DMA range outside IOMMU capability\n");
 			return -EFAULT;
 		}
-		/* ...then finally give it a kicking to make sure it fits */
-		base_pfn = max(base, domain->geometry.aperture_start) >> order;
 	}
+	/* ...then finally give it a kicking to make sure it fits */
+	base_pfn = max_t(unsigned long, base_pfn,
+			 domain->geometry.aperture_start >> order);
 
 	/* start_pfn is always nonzero for an already-initialised domain */
 	mutex_lock(&cookie->mutex);
-- 
2.39.2.101.g768bb238c484.dirty

On Mon, May 20, 2024 at 08:14:44PM GMT, Robin Murphy wrote:
> Despite carefully rewording the kerneldoc to describe the new direct
> interaction with dma_range_map, it seems I managed to confuse myself in
> removing the redundant force_aperture check and ended up making the code
> not do that at all. This led to dma_range_maps inadvertently being able
> to set iovad->start_pfn = 0, and all the nonsensical chaos which ensues
> from there. Restore the correct behaviour of constraining base_pfn to
> the domain aperture regardless of dma_range_map, and not trying to apply
> dma_range_map constraints to the basic IOVA domain since they will be
> properly handled with reserved regions later.
> 
> Reported-by: Jon Hunter <jonathanh@nvidia.com>
> Reported-by: Jerry Snitselaar <jsnitsel@redhat.com>
> Fixes: e28a114243d1 ("iommu/dma: Make limit checks self-contained")

I'm seeing this upstream as ad4750b07d34 ("iommu/dma: Make limit checks self-contained") now.

> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
> ---

Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
Reviewed-by: Jerry Snitselaar <jsnitsel@redhat.com>

>  drivers/iommu/dma-iommu.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> index 89a53c2f2cf9..1ec89ca75c59 100644
> --- a/drivers/iommu/dma-iommu.c
> +++ b/drivers/iommu/dma-iommu.c
> @@ -685,15 +685,15 @@ static int iommu_dma_init_domain(struct iommu_domain *domain, struct device *dev
>  
>  	/* Check the domain allows at least some access to the device... */
>  	if (map) {
> -		dma_addr_t base = dma_range_map_min(map);
> -		if (base > domain->geometry.aperture_end ||
> +		if (dma_range_map_min(map) > domain->geometry.aperture_end ||
>  		    dma_range_map_max(map) < domain->geometry.aperture_start) {
>  			pr_warn("specified DMA range outside IOMMU capability\n");
>  			return -EFAULT;
>  		}
> -		/* ...then finally give it a kicking to make sure it fits */
> -		base_pfn = max(base, domain->geometry.aperture_start) >> order;
>  	}
> +	/* ...then finally give it a kicking to make sure it fits */
> +	base_pfn = max_t(unsigned long, base_pfn,
> +			 domain->geometry.aperture_start >> order);
>  
>  	/* start_pfn is always nonzero for an already-initialised domain */
>  	mutex_lock(&cookie->mutex);
> -- 
> 2.39.2.101.g768bb238c484.dirty
>

On 21/05/2024 5:45 am, Jerry Snitselaar wrote:
> On Mon, May 20, 2024 at 08:14:44PM GMT, Robin Murphy wrote:
>> Despite carefully rewording the kerneldoc to describe the new direct
>> interaction with dma_range_map, it seems I managed to confuse myself in
>> removing the redundant force_aperture check and ended up making the code
>> not do that at all. This led to dma_range_maps inadvertently being able
>> to set iovad->start_pfn = 0, and all the nonsensical chaos which ensues
>> from there. Restore the correct behaviour of constraining base_pfn to
>> the domain aperture regardless of dma_range_map, and not trying to apply
>> dma_range_map constraints to the basic IOVA domain since they will be
>> properly handled with reserved regions later.
>>
>> Reported-by: Jon Hunter <jonathanh@nvidia.com>
>> Reported-by: Jerry Snitselaar <jsnitsel@redhat.com>
>> Fixes: e28a114243d1 ("iommu/dma: Make limit checks self-contained")
> 
> I'm seeing this upstream as ad4750b07d34 ("iommu/dma: Make limit checks self-contained") now.

What, you mean when Linus pulls some patches I wrote he *doesn't* also 
rebase my local development branch for me? /facepalm

Joerg, please shout if you'd like a resend for that.

>> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
>> ---
> 
> Tested-by: Jerry Snitselaar <jsnitsel@redhat.com>
> Reviewed-by: Jerry Snitselaar <jsnitsel@redhat.com>

Thanks!
Robin.

>>   drivers/iommu/dma-iommu.c | 8 ++++----
>>   1 file changed, 4 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>> index 89a53c2f2cf9..1ec89ca75c59 100644
>> --- a/drivers/iommu/dma-iommu.c
>> +++ b/drivers/iommu/dma-iommu.c
>> @@ -685,15 +685,15 @@ static int iommu_dma_init_domain(struct iommu_domain *domain, struct device *dev
>>   
>>   	/* Check the domain allows at least some access to the device... */
>>   	if (map) {
>> -		dma_addr_t base = dma_range_map_min(map);
>> -		if (base > domain->geometry.aperture_end ||
>> +		if (dma_range_map_min(map) > domain->geometry.aperture_end ||
>>   		    dma_range_map_max(map) < domain->geometry.aperture_start) {
>>   			pr_warn("specified DMA range outside IOMMU capability\n");
>>   			return -EFAULT;
>>   		}
>> -		/* ...then finally give it a kicking to make sure it fits */
>> -		base_pfn = max(base, domain->geometry.aperture_start) >> order;
>>   	}
>> +	/* ...then finally give it a kicking to make sure it fits */
>> +	base_pfn = max_t(unsigned long, base_pfn,
>> +			 domain->geometry.aperture_start >> order);
>>   
>>   	/* start_pfn is always nonzero for an already-initialised domain */
>>   	mutex_lock(&cookie->mutex);
>> -- 
>> 2.39.2.101.g768bb238c484.dirty
>>
>

On Tue, May 21, 2024 at 04:14:50PM +0100, Robin Murphy wrote:
> Joerg, please shout if you'd like a resend for that.

No need to re-send. I applied with the correct commit-id.