Re: [RFC V2] iommu: correct group reference count

[Will Deacon <will.deacon@arm.com>]

On Mon, Nov 09, 2015 at 02:13:28PM +0800, Peng Fan wrote:
> The basic flow for iommu_group_for_dev is:
> iommu_group_get_for_dev
>         |->  iommu_group_get   : increase reference count by 1.
>              return group;
>         |->  ops->device_group : Init/increase reference count to/by 1.
>              iommu_group_add_device  : Increase reference count by 1.
>              return group;
> 
> We can see that ops->device_group and iommu_group_add_device will together
> increase the iommu group reference count by 2. Actually we only need 1,
> but not 2. So we need add iommu_group_put after iommu_group_add_device
> to make sure iommu_group_get_for_dev only increase reference count by 1.
> 
> Signed-off-by: Peng Fan <van.freenix@gmail.com>
> Cc: Joerg Roedel <joro@8bytes.org>
> Cc: Will Deacon <will.deacon@arm.com>
> ---
> 
> V1 thread: https://lkml.org/lkml/2015/11/3/304
> Changes V2:
>  I did not see the update about device_group when I worked out V1. So
>  redo the patch and refine commit msg and rebased to latest linus'
>  linux master tree.
> 
>  drivers/iommu/iommu.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index abae363..9c1971b 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -852,10 +852,10 @@ struct iommu_group *iommu_group_get_for_dev(struct device *dev)
>  	}
>  
>  	ret = iommu_group_add_device(group, dev);
> -	if (ret) {
> -		iommu_group_put(group);
> +	iommu_group_put(group);
> +
> +	if (ret)
>  		return ERR_PTR(ret);
> -	}

Hmm, I don't think this is correct either:

  (1) It doesn't help the arm-smmu driver (which deals with platform
      devices too)

  (2) It breaks IOMMU drivers that currently have the explicit put already
      (e.g. amd-iommu.c:init_iommu_group)

The easiest way for you to proceed is to extend your original patch so
that it also adds an iommu_group_put to the arm-smmu-v3 driver after
a successful iommu_group_get_for_dev.

Will