[v2,5/8] iommu: Switch __iommu_domain_alloc() to device ops

  In all the places we allocate default domains, we have (or can easily
get hold of) a device from which to resolve the right IOMMU ops; only
the public iommu_domain_alloc() interface actually depends on bus ops.
Reworking the public API is a big enough mission in its own right, but
in the meantime we can still decouple it from bus ops internally to move
forward.

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

v2: - Explain the mitigation better in the warning message
    - Fix theoretical bug if alloc_dev is never assigned because the
      bus has no devices
    - Use new dev_iommu_ops_valid() since in theory VFIO noiommu makes
      device_iommu_mapped() -> dev_iommu_ops() an unsafe assumption
      [Baolu]

 drivers/iommu/iommu.c | 59 ++++++++++++++++++++++++++++---------------
 1 file changed, 39 insertions(+), 20 deletions(-)
  

Hi Robin,

On Thu, 26 Jan 2023 18:26:20 +0000, Robin Murphy <robin.murphy@arm.com>
wrote:

>  
> +static int __iommu_domain_alloc_dev(struct device *dev, void *data)
> +{
> +	struct device **alloc_dev = data;
> +
> +	if (!dev_iommu_ops_valid(dev))
> +		return 0;
> +
> +	WARN_ONCE(*alloc_dev && dev_iommu_ops(dev) !=
> dev_iommu_ops(*alloc_dev),
> +		"Multiple IOMMU drivers present, which the public IOMMU
> API can't fully support yet. You may still need to disable one or more to
> get the expected result here, sorry!\n"); +
> +	*alloc_dev = dev;
> +	return 0;
> +}
> +
>  struct iommu_domain *iommu_domain_alloc(struct bus_type *bus)
>  {
> -	return __iommu_domain_alloc(bus, IOMMU_DOMAIN_UNMANAGED);
> +	struct device *dev = NULL;
> +
> +	/* We always check the whole bus, so the return value isn't
> useful */
> +	bus_for_each_dev(bus, NULL, &dev, __iommu_domain_alloc_dev);
> +	if (!dev)
> +		return NULL;
Since __iommu_domain_alloc_dev() will always return 0, bus_for_each_dev()
will never breakout until the whole dev list is iterated over. If so, would
dev only record the last one? i.e. prior results get overwritten.  Maybe a
misunderstood the logic.

> +	return __iommu_domain_alloc(dev, IOMMU_DOMAIN_UNMANAGED);
>  }
>  EXPORT_SYMBOL_GPL(iommu_domain_alloc);

Thanks,

Jacob
  

On 2023-01-26 23:22, Jacob Pan wrote:
> Hi Robin,
> 
> On Thu, 26 Jan 2023 18:26:20 +0000, Robin Murphy <robin.murphy@arm.com>
> wrote:
> 
>>   
>> +static int __iommu_domain_alloc_dev(struct device *dev, void *data)
>> +{
>> +	struct device **alloc_dev = data;
>> +
>> +	if (!dev_iommu_ops_valid(dev))
>> +		return 0;
>> +
>> +	WARN_ONCE(*alloc_dev && dev_iommu_ops(dev) !=
>> dev_iommu_ops(*alloc_dev),
>> +		"Multiple IOMMU drivers present, which the public IOMMU
>> API can't fully support yet. You may still need to disable one or more to
>> get the expected result here, sorry!\n"); +
>> +	*alloc_dev = dev;
>> +	return 0;
>> +}
>> +
>>   struct iommu_domain *iommu_domain_alloc(struct bus_type *bus)
>>   {
>> -	return __iommu_domain_alloc(bus, IOMMU_DOMAIN_UNMANAGED);
>> +	struct device *dev = NULL;
>> +
>> +	/* We always check the whole bus, so the return value isn't
>> useful */
>> +	bus_for_each_dev(bus, NULL, &dev, __iommu_domain_alloc_dev);
>> +	if (!dev)
>> +		return NULL;
> Since __iommu_domain_alloc_dev() will always return 0, bus_for_each_dev()
> will never breakout until the whole dev list is iterated over. If so, would
> dev only record the last one? i.e. prior results get overwritten.  Maybe a
> misunderstood the logic.

Yes, as the comment points out, the intent is to walk the whole bus to 
check it for consistency. Beyond that, we just need *a* device with 
IOMMU ops; it doesn't matter at all which one it is. It happens to be 
the last one off the list because that's what fell out of writing the 
fewest lines of code.

(You could argue that there's no need to repeat the full walk if the 
WARN_ONCE has already fired, but I'd rather keep the behaviour simple 
and consistent - this is only meant to be a short-term solution, and 
it's not a performance-critical path)

Thanks,
Robin.

> 
>> +	return __iommu_domain_alloc(dev, IOMMU_DOMAIN_UNMANAGED);
>>   }
>>   EXPORT_SYMBOL_GPL(iommu_domain_alloc);
> 
> Thanks,
> 
> Jacob
  

Hi Robin,

On Fri, 27 Jan 2023 11:42:27 +0000, Robin Murphy <robin.murphy@arm.com>
wrote:

> On 2023-01-26 23:22, Jacob Pan wrote:
> > Hi Robin,
> > 
> > On Thu, 26 Jan 2023 18:26:20 +0000, Robin Murphy <robin.murphy@arm.com>
> > wrote:
> >   
> >>   
> >> +static int __iommu_domain_alloc_dev(struct device *dev, void *data)
> >> +{
> >> +	struct device **alloc_dev = data;
> >> +
> >> +	if (!dev_iommu_ops_valid(dev))
> >> +		return 0;
> >> +
> >> +	WARN_ONCE(*alloc_dev && dev_iommu_ops(dev) !=
> >> dev_iommu_ops(*alloc_dev),
> >> +		"Multiple IOMMU drivers present, which the public
> >> IOMMU API can't fully support yet. You may still need to disable one
> >> or more to get the expected result here, sorry!\n"); +
> >> +	*alloc_dev = dev;
> >> +	return 0;
> >> +}
> >> +
> >>   struct iommu_domain *iommu_domain_alloc(struct bus_type *bus)
> >>   {
> >> -	return __iommu_domain_alloc(bus, IOMMU_DOMAIN_UNMANAGED);
> >> +	struct device *dev = NULL;
> >> +
> >> +	/* We always check the whole bus, so the return value isn't
> >> useful */
> >> +	bus_for_each_dev(bus, NULL, &dev, __iommu_domain_alloc_dev);
> >> +	if (!dev)
> >> +		return NULL;  
> > Since __iommu_domain_alloc_dev() will always return 0,
> > bus_for_each_dev() will never breakout until the whole dev list is
> > iterated over. If so, would dev only record the last one? i.e. prior
> > results get overwritten.  Maybe a misunderstood the logic.  
> 
> Yes, as the comment points out, the intent is to walk the whole bus to 
> check it for consistency. Beyond that, we just need *a* device with 
> IOMMU ops; it doesn't matter at all which one it is. It happens to be 
> the last one off the list because that's what fell out of writing the 
> fewest lines of code.
> 
> (You could argue that there's no need to repeat the full walk if the 
> WARN_ONCE has already fired, but I'd rather keep the behaviour simple 
> and consistent - this is only meant to be a short-term solution, and 
> it's not a performance-critical path)
That make sense now, thank you for the explanation.


Jacob
  

On 2023/1/27 2:26, Robin Murphy wrote:
> In all the places we allocate default domains, we have (or can easily
> get hold of) a device from which to resolve the right IOMMU ops; only
> the public iommu_domain_alloc() interface actually depends on bus ops.
> Reworking the public API is a big enough mission in its own right, but
> in the meantime we can still decouple it from bus ops internally to move
> forward.
> 
> Signed-off-by: Robin Murphy<robin.murphy@arm.com>

Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>

Best regards,
baolu
  

On Thu, Jan 26, 2023 at 06:26:20PM +0000, Robin Murphy wrote:

> @@ -1980,9 +1978,30 @@ static struct iommu_domain *__iommu_domain_alloc(struct bus_type *bus,
>  	return domain;
>  }
>  
> +static int __iommu_domain_alloc_dev(struct device *dev, void *data)
> +{
> +	struct device **alloc_dev = data;
> +
> +	if (!dev_iommu_ops_valid(dev))
> +		return 0;

This valid check is necessary to be done in __iommu_domain_alloc()
because the callback coming from iommu_device_claim_dma_owner() never
validated that the device/group has an iommu driver:

> @@ -3120,13 +3139,13 @@ static int __iommu_group_alloc_blocking_domain(struct iommu_group *group)
>  	if (group->blocking_domain)
>  		return 0;
>  
> -	group->blocking_domain = __iommu_domain_alloc(dev->bus, IOMMU_DOMAIN_BLOCKED);
> +	group->blocking_domain = __iommu_domain_alloc(dev, IOMMU_DOMAIN_BLOCKED);

Jason