The IOMMU group sysfs interface allows users to change the default
domain of a group. The current implementation uses device_lock() to make
sure that the devices in the group are not bound to any driver and won't
be bound during the process of changing the default domain. In order to
avoid a possible deadlock caused by lock order of device_lock and
group->mutex, it limits the functionality to singleton groups only.

The recently implemented DMA ownership framework can be applied here to
replace device_lock(). In addition, use group->mutex to ensure that the
iommu ops of the devices is always valid during the process of changing
default domain.

With above replacement and enhancement, the device_lock() could be
removed and the singleton-group-only limitation could be removed.

The whole series is also available on github:
https://github.com/LuBaolu/intel-iommu/commits/iommu-sysfs-default-domain-extension-v2

This series is on top of below series from Robin,
https://lore.kernel.org/linux-iommu/cover.1674753627.git.robin.murphy@arm.com/

Please help to review and suggest.

Change log:
v2:
 - Use group->mutex instead of an additional rw lock.

v1: initial post
 - https://lore.kernel.org/linux-iommu/20230213074941.919324-1-baolu.lu@linux.intel.com/

Lu Baolu (6):
  ARM/dma-mapping: Remove iommu_detach_device()
  iommu: Split iommu_group_remove_device() into helpers
  iommu: Same critical region for device release and removal
  iommu: Move lock from iommu_change_dev_def_domain() to its caller
  iommu: Replace device_lock() with group->mutex
  iommu: Cleanup iommu_change_dev_def_domain()

 arch/arm/mm/dma-mapping.c                     |   1 -
 drivers/iommu/iommu.c                         | 255 +++++++-----------
 .../ABI/testing/sysfs-kernel-iommu_groups     |   1 -
 3 files changed, 102 insertions(+), 155 deletions(-)

-- 
2.34.1

On Fri, Feb 17, 2023 at 05:47:30PM +0800, Lu Baolu wrote:
> The IOMMU group sysfs interface allows users to change the default
> domain of a group. The current implementation uses device_lock() to make
> sure that the devices in the group are not bound to any driver and won't
> be bound during the process of changing the default domain. In order to
> avoid a possible deadlock caused by lock order of device_lock and
> group->mutex, it limits the functionality to singleton groups only.
> 
> The recently implemented DMA ownership framework can be applied here to
> replace device_lock(). In addition, use group->mutex to ensure that the
> iommu ops of the devices is always valid during the process of changing
> default domain.
> 
> With above replacement and enhancement, the device_lock() could be
> removed and the singleton-group-only limitation could be removed.
> 
> The whole series is also available on github:
> https://github.com/LuBaolu/intel-iommu/commits/iommu-sysfs-default-domain-extension-v2
> 
> This series is on top of below series from Robin,
> https://lore.kernel.org/linux-iommu/cover.1674753627.git.robin.murphy@arm.com/
> 
> Please help to review and suggest.

Given the overall situation, I think my suggestion to use
arm_iommu_release_device() might be more short term practical.

Jason

On 2/17/23 11:47 PM, Jason Gunthorpe wrote:
> On Fri, Feb 17, 2023 at 05:47:30PM +0800, Lu Baolu wrote:
>> The IOMMU group sysfs interface allows users to change the default
>> domain of a group. The current implementation uses device_lock() to make
>> sure that the devices in the group are not bound to any driver and won't
>> be bound during the process of changing the default domain. In order to
>> avoid a possible deadlock caused by lock order of device_lock and
>> group->mutex, it limits the functionality to singleton groups only.
>>
>> The recently implemented DMA ownership framework can be applied here to
>> replace device_lock(). In addition, use group->mutex to ensure that the
>> iommu ops of the devices is always valid during the process of changing
>> default domain.
>>
>> With above replacement and enhancement, the device_lock() could be
>> removed and the singleton-group-only limitation could be removed.
>>
>> The whole series is also available on github:
>> https://github.com/LuBaolu/intel-iommu/commits/iommu-sysfs-default-domain-extension-v2
>>
>> This series is on top of below series from Robin,
>> https://lore.kernel.org/linux-iommu/cover.1674753627.git.robin.murphy@arm.com/
>>
>> Please help to review and suggest.
> Given the overall situation, I think my suggestion to use
> arm_iommu_release_device() might be more short term practical.

Yes. Fair enough.

Best regards,
baolu