Re: [PATCH 2/2] iommu/arm-smmu: Allow client devices to select direct mapping

[Robin Murphy]

On 2020-04-16 5:23 pm, Sai Prakash Ranjan wrote:
> Hi Robin,
>
> On 2020-04-16 19:28, Robin Murphy wrote:
> > On 2020-01-22 11:48 am, Sai Prakash Ranjan wrote:
> > > From: Jordan Crouse <jcrouse@xxxxxxxxxxxxxx>
> > >
> > > Some client devices want to directly map the IOMMU themselves instead
> > > of using the DMA domain. Allow those devices to opt in to direct
> > > mapping by way of a list of compatible strings.
> > >
> > > Signed-off-by: Jordan Crouse <jcrouse@xxxxxxxxxxxxxx>
> > > Co-developed-by: Sai Prakash Ranjan <saiprakash.ranjan@xxxxxxxxxxxxxx>
> > > Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@xxxxxxxxxxxxxx>
> > > ---
> > > Â drivers/iommu/arm-smmu-qcom.c | 39 +++++++++++++++++++++++++++++++++++
> > > Â drivers/iommu/arm-smmu.cÂÂÂÂÂ |Â 3 +++
> > > Â drivers/iommu/arm-smmu.hÂÂÂÂÂ |Â 5 +++++
> > > Â 3 files changed, 47 insertions(+)
> > >
> > > diff --git a/drivers/iommu/arm-smmu-qcom.c 
> > > b/drivers/iommu/arm-smmu-qcom.c
> > > index 64a4ab270ab7..ff746acd1c81 100644
> > > --- a/drivers/iommu/arm-smmu-qcom.c
> > > +++ b/drivers/iommu/arm-smmu-qcom.c
> > > @@ -3,6 +3,7 @@
> > > ÂÂ * Copyright (c) 2019, The Linux Foundation. All rights reserved.
> > > ÂÂ */
> > > Â +#include <linux/of_device.h>
> > > Â #include <linux/qcom_scm.h>
> > > ÂÂÂ #include "arm-smmu.h"
> > > @@ -11,6 +12,43 @@ struct qcom_smmu {
> > > ÂÂÂÂÂ struct arm_smmu_device smmu;
> > > Â };
> > > Â +static const struct arm_smmu_client_match_data qcom_adreno = {
> > > +ÂÂÂ .direct_mapping = true,
> > > +};
> > > +
> > > +static const struct arm_smmu_client_match_data qcom_mdss = {
> > > +ÂÂÂ .direct_mapping = true,
> > > +};
> >
> > Might it make sense to group these by the desired SMMU behaviour
> > rather than (apparently) what kind of device the client happens to be,
> > which seems like a completely arbitrary distinction from the SMMU
> > driver's PoV?
>
> Sorry, I did not get the "grouping by the desired SMMU behaviour" thing.
> Could you please give some more details?

I mean this pattern:

device_a_data {
	.thing = this;
}

device_b_data {
	.thing = this;
}

device_c_data {
	.thing = that;
}

match[] = {
	{ .compatible = "A", .data = &device_a_data },
	{ .compatible = "B", .data = &device_b_data },
	{ .compatible = "C", .data = &device_c_data },
};

...vs. this pattern:

do_this {
	.thing = this;
}

do_that {
	.thing = that;
}

match[] = {
	{ .compatible = "A", .data = &do_this },
	{ .compatible = "B", .data = &do_this },
	{ .compatible = "C", .data = &do_that },
};

From the perspective of the thing doing the thing, grouping the data by 
device is meaningless if all that matters is whether to do this or that. 
The second pattern expresses that more naturally.

Of course if every device turns out to need a unique combination of 
several behaviours, then there ends up being no practical difference 
except that it's probably easier to come up with nice names under the 
first pattern. I guess it's up to how you see this developing in future; 
whether you're likely to need fine-grained per-device control, or don't 
expect it to go much beyond domain type.

Robin.