Re: [PATCH 1/2] ARM: SMMU: add devices attached to the SMMU to anIOMMU group

From: Antonios Motakis
Date: Fri Oct 18 2013 - 06:30:11 EST


On Mon, Oct 14, 2013 at 7:07 PM, Will Deacon <will.deacon@xxxxxxx> wrote:
> On Mon, Oct 14, 2013 at 04:13:15PM +0100, Antonios Motakis wrote:
>> On Mon, Oct 14, 2013 at 2:48 PM, Will Deacon <will.deacon@xxxxxxx> wrote:
>> > > diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
>> > > index 0f45a48..8b71332 100644
>> > > --- a/drivers/iommu/arm-smmu.c
>> > > +++ b/drivers/iommu/arm-smmu.c
>> > > @@ -1502,6 +1502,8 @@ static int arm_smmu_add_device(struct device *dev)
>> > > {
>> > > struct arm_smmu_device *child, *parent, *smmu;
>> > > struct arm_smmu_master *master = NULL;
>> > > + struct iommu_group *group;
>> > > + int ret;
>> > >
>> > > spin_lock(&arm_smmu_devices_lock);
>> > > list_for_each_entry(parent, &arm_smmu_devices, list) {
>> > > @@ -1534,13 +1536,27 @@ static int arm_smmu_add_device(struct device *dev)
>> > > if (!master)
>> > > return -ENODEV;
>> > >
>> > > + group = iommu_group_get(dev);
>> >
>> > I'm not especially familiar with IOMMU groups (I understand them as the
>> > minimum translation granularity, which would mean single StreamID for the
>> > ARM SMMU), but under what circumstances would you expect to receive a
>> > non-NULL group here? I can't see any other code adding devices to groups
>> > (outside of other drivers)...
>> >
>>
>> You are right, only other IOMMU drivers will add a device to a group.
>> There was a discussion about this when I posted a similar patch for
>> the Exynos System MMU driver, see
>> http://lists.infradead.org/pipermail/linux-arm-kernel/2013-July/185675.html
>>
>> The idea is to check in the case of add_device() being called multiple
>> times, which is not the case most of the time, but still a sane
>> safeguard.
>
> Ok, but it feels a bit weird. The current code (arm_smmu_add_device)
> basically does a bunch of sanity checking against the DT data in order to
> find where the master sits in the device topology. Then it updates
> dev->archdata.iommu to point at the relevant SMMU instance.
>
> So, the interesting case is where the device was previously associated with
> a *different* IOMMU. In that case, the current code clobbers the iommu field
> with the new smmu, whereas the new code could end up getting very confused
> with respect to IOMMU groups.
>
> A better way is probably to check that dev->archdata.iommu is NULL before we
> assign to it. If not, then spit out a warning and return an error. That
> would also mean you could get rid of the group get/put calls.
>

Good point, this is a better way to handle this. I'll respin based on this.

> Will
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/