iommu_dma_get_resv_regions() assumes that iommu_fwspec field for
corresponding device is set which is not always true. Since
iommu_dma_get_resv_regions() seems to be a future-proof generic API
that can be used by any iommu driver, add an explicit check for NULL.
Currently it can work by accident since compiler can eliminate
the 'iommu_fwspec' check altogether when CONFIG_ACPI_IORT=n, but
code elimination from optimizations is not reliable.
Signed-off-by: Aleksandr Fedorov <[email protected]>
---
A compilation failure has been observed on a gcc-compatible compiler based on EDG.
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index d85d54f2b549..474b1b7211d7 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -382,10 +382,10 @@ void iommu_put_dma_cookie(struct iommu_domain *domain)
*/
void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
{
+ struct iommu_fwspec *iommu_fwspec = dev_iommu_fwspec_get(dev);
- if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
+ if (iommu_fwspec && !is_of_node(iommu_fwspec->iommu_fwnode))
iort_iommu_msi_get_resv_regions(dev, list);
-
}
EXPORT_SYMBOL(iommu_dma_get_resv_regions);
On 2022-02-09 14:09, Aleksandr Fedorov wrote:
> iommu_dma_get_resv_regions() assumes that iommu_fwspec field for
> corresponding device is set which is not always true. Since
> iommu_dma_get_resv_regions() seems to be a future-proof generic API
> that can be used by any iommu driver, add an explicit check for NULL.
Except it's not a "generic" interface for drivers to call at random,
it's a helper for retrieving common firmware-based information
specifically for drivers already using the fwspec mechanism for common
firmware bindings. If any driver calls this with a device *without* a
valid fwnode, it deserves to crash because it's done something
fundamentally wrong.
I concur that it's not exactly obvious that "non-IOMMU-specific" means
"based on common firmware bindings, thus implying fwspec".
Robin.
> Currently it can work by accident since compiler can eliminate
> the 'iommu_fwspec' check altogether when CONFIG_ACPI_IORT=n, but
> code elimination from optimizations is not reliable.
>
> Signed-off-by: Aleksandr Fedorov <[email protected]>
> ---
> A compilation failure has been observed on a gcc-compatible compiler based on EDG.
>
> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> index d85d54f2b549..474b1b7211d7 100644
> --- a/drivers/iommu/dma-iommu.c
> +++ b/drivers/iommu/dma-iommu.c
> @@ -382,10 +382,10 @@ void iommu_put_dma_cookie(struct iommu_domain *domain)
> */
> void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
> {
> + struct iommu_fwspec *iommu_fwspec = dev_iommu_fwspec_get(dev);
>
> - if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
> + if (iommu_fwspec && !is_of_node(iommu_fwspec->iommu_fwnode))
> iort_iommu_msi_get_resv_regions(dev, list);
> -
> }
> EXPORT_SYMBOL(iommu_dma_get_resv_regions);
>
> _______________________________________________
> iommu mailing list
> [email protected]
> https://lists.linuxfoundation.org/mailman/listinfo/iommu
> On 2022-02-09 14:09, Aleksandr Fedorov wrote:
>> iommu_dma_get_resv_regions() assumes that iommu_fwspec field for
>> corresponding device is set which is not always true. Since
>> iommu_dma_get_resv_regions() seems to be a future-proof generic API
>> that can be used by any iommu driver, add an explicit check for NULL.
>
> Except it's not a "generic" interface for drivers to call at random,
> it's a helper for retrieving common firmware-based information
> specifically for drivers already using the fwspec mechanism for common
> firmware bindings. If any driver calls this with a device *without* a
> valid fwnode, it deserves to crash because it's done something
> fundamentally wrong.
>
> I concur that it's not exactly obvious that "non-IOMMU-specific" means
> "based on common firmware bindings, thus implying fwspec".
Thanks for the explanations, yes, this was the misunderstanding on my
part. Maybe add a comment?
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index d85d54f2b549..ce5e7d4d054a 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -379,6 +379,9 @@ void iommu_put_dma_cookie(struct iommu_domain *domain)
* for general non-IOMMU-specific reservations. Currently, this covers GICv3
* ITS region reservation on ACPI based ARM platforms that may require HW MSI
* reservation.
+ *
+ * Note that this helper is meant to be used only by drivers that are already
+ * using the fwspec mechanism for common firmware bindings.
*/
void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
{