On Mon, Jul 25, 2016 at 04:51:00PM +0100, Robin Murphy wrote:

On 25/07/16 16:41, Lorenzo Pieralisi wrote:
[...]

quoted

quoted

quoted

diff --git a/include/linux/of_iommu.h b/include/linux/of_iommu.h
index 308791f..2362232 100644
--- a/include/linux/of_iommu.h
+++ b/include/linux/of_iommu.h

@@ -15,13 +15,8 @@ extern void of_iommu_init(void);
 extern const struct iommu_ops *of_iommu_configure(struct device *dev,
 					struct device_node *master_np);
 
-struct iommu_fwspec {
-	const struct iommu_ops	*iommu_ops;
-	struct device_node	*iommu_np;
-	void			*iommu_priv;
-	unsigned int		num_ids;
-	u32			ids[];
-};
+void of_iommu_set_ops(struct device_node *np, const struct iommu_ops *ops);
+const struct iommu_ops *of_iommu_get_ops(struct device_node *np);

Is there some reason we need to retain the existing definitions of
these? I was assuming we'd be able to move the entire implementation
over to the fwspec code and leave behind nothing more than trivial
wrappers, e.g.:

#define of_iommu_get_ops(np) iommu_fwspec_get_ops(&(np)->fwnode_handle)

Yep, that's exactly what I did but then I was bitten by config
dependencies. If we implement of_iommu_get/set_ops() as wrappers,
we have to compile iommu_fwspec_get/set_ops() on arches that may
not have struct dev_archdata.iommu, unless we introduce yet another
config symbol to avoid compiling that code (see eg iommu_fwspec_init(),
we can't compile it on eg x86 even though we do need of_iommu_get_ops()
on it - so iommu_fwspec_get_ops(), that lives in the same compilation
unit as eg iommu_fwspec_init()).

So short answer is: there is no reason apart from dev_archdata.iommu
being arch specific, if we were able to move iommu_fwspec to generic
code (ie struct device, somehow) I would certainly get rid of this
stupid code duplication (or as I said I can add a config entry for
that, more ideas are welcome).

OK, given Rob's comment as well, I guess breaking that dependency is to
everyone's benefit. Since it's quite closely related, how about if we
follow the arch_setup_dma_ops() pattern with an
arch_{get,set}_iommu_fwspec(dev) type thing?

Yes we can do that too as an intermediate step, that solves the
problem (and it makes this patch much simpler), it is cleaner
than doing it with a(nother) Kconfig entry.

Thanks,
Lorenzo

Robin.

quoted

Thanks,
Lorenzo

quoted

Robin.

quoted

 #else
 

@@ -39,17 +34,14 @@ static inline const struct iommu_ops *of_iommu_configure(struct device *dev,
 	return NULL;
 }
 
-struct iommu_fwspec;
-
-#endif	/* CONFIG_OF_IOMMU */
+static inline void of_iommu_set_ops(struct device_node *np,
+				    const struct iommu_ops *ops)
+{ }
 
-int iommu_fwspec_init(struct device *dev, struct device_node *iommu_np);
-void iommu_fwspec_free(struct device *dev);
-int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids);
-struct iommu_fwspec *dev_iommu_fwspec(struct device *dev);
+static inline const struct iommu_ops *
+of_iommu_get_ops(struct device_node *np) { return NULL; }
 
-void of_iommu_set_ops(struct device_node *np, const struct iommu_ops *ops);
-const struct iommu_ops *of_iommu_get_ops(struct device_node *np);
+#endif	/* CONFIG_OF_IOMMU */
 
 extern struct of_device_id __iommu_of_table;