Linux Virtualization - Mar 2020 - [PATCH 00/14] iommu: Move iommu_fwspec out of 'struct device'

On Fri, Mar 06, 2020 at 04:39:37PM +0800, Hanjun Guo
wrote:
> Hi Joerg,
> 
> On 2020/2/28 23:08, Joerg Roedel wrote:
> > Hi,
> > 
> > here is a patch-set to rename iommu_param to dev_iommu and
> > establish it as a struct for generic per-device iommu-data.
> > Also move the iommu_fwspec pointer from struct device into
> > dev_iommu to have less iommu-related pointers in struct
> > device.
> > 
> > The bigger part of this patch-set moves the iommu_priv
> > pointer from struct iommu_fwspec to dev_iommu, making is
> > usable for iommu-drivers which do not use fwspecs.
> > 
> > The changes for that were mostly straightforward, except for
> > the arm-smmu (_not_ arm-smmu-v3) and the qcom iommu driver.
> > Unfortunatly I don't have the hardware for those, so any
> > testing of these drivers is greatly appreciated.
> 
> I tested this patch set on Kunpeng 920 ARM64 server which
> using smmu-v3 with ACPI booting, but triggered a NULL
> pointer dereference and panic at boot:
I think that's because patch 01/14 move the fwspec access too early. In 

                err = pci_for_each_dma_alias(to_pci_dev(dev),
                                             iort_pci_iommu_init, &info);

                if (!err && iort_pci_rc_supports_ats(node))
                        dev->iommu_fwspec->flags |=
IOMMU_FWSPEC_PCI_RC_ATS;

the iommu_fwspec is only valid if iort_pci_iommu_init() initialized it
successfully, if err == 0. The following might fix it:

diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c
index 0e981d7f3c7d..7d04424189df 100644
--- a/drivers/acpi/arm64/iort.c
+++ b/drivers/acpi/arm64/iort.c
@@ -1015,7 +1015,7 @@ const struct iommu_ops *iort_iommu_configure(struct device
*dev)
 		return ops;

 	if (dev_is_pci(dev)) {
-		struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
+		struct iommu_fwspec *fwspec;
 		struct pci_bus *bus = to_pci_dev(dev)->bus;
 		struct iort_pci_alias_info info = { .dev = dev };

@@ -1028,7 +1028,8 @@ const struct iommu_ops *iort_iommu_configure(struct device
*dev)
 		err = pci_for_each_dma_alias(to_pci_dev(dev),
 					     iort_pci_iommu_init, &info);

-		if (!err && iort_pci_rc_supports_ats(node))
+		fwspec = dev_iommu_fwspec_get(dev);
+		if (fwspec && iort_pci_rc_supports_ats(node))
 			fwspec->flags |= IOMMU_FWSPEC_PCI_RC_ATS;
 	} else {
 		int i = 0;


Note that this use of iommu_fwspec will be removed by the ATS cleanup
series [1], but this change should work as a temporary fix.

Thanks,
Jean

[1] https://lore.kernel.org/linux-iommu/20200213165049.508908-10-jean-philippe
at linaro.org/

Hi Jean-Philippe,

On Fri, Mar 06, 2020 at 11:09:55AM +0100, Jean-Philippe Brucker
wrote:
> I think that's because patch 01/14 move the fwspec access too early. In
> 
>                 err = pci_for_each_dma_alias(to_pci_dev(dev),
>                                              iort_pci_iommu_init,
&info);
> 
>                 if (!err && iort_pci_rc_supports_ats(node))
>                         dev->iommu_fwspec->flags |=
IOMMU_FWSPEC_PCI_RC_ATS;
> 
> the iommu_fwspec is only valid if iort_pci_iommu_init() initialized it
> successfully, if err == 0. The following might fix it:
> 
> diff --git a/drivers/acpi/arm64/iort.c b/drivers/acpi/arm64/iort.c
> index 0e981d7f3c7d..7d04424189df 100644
> --- a/drivers/acpi/arm64/iort.c
> +++ b/drivers/acpi/arm64/iort.c
> @@ -1015,7 +1015,7 @@ const struct iommu_ops *iort_iommu_configure(struct
device *dev)
>  		return ops;
> 
>  	if (dev_is_pci(dev)) {
> -		struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
> +		struct iommu_fwspec *fwspec;
>  		struct pci_bus *bus = to_pci_dev(dev)->bus;
>  		struct iort_pci_alias_info info = { .dev = dev };
> 
> @@ -1028,7 +1028,8 @@ const struct iommu_ops *iort_iommu_configure(struct
device *dev)
>  		err = pci_for_each_dma_alias(to_pci_dev(dev),
>  					     iort_pci_iommu_init, &info);
> 
> -		if (!err && iort_pci_rc_supports_ats(node))
> +		fwspec = dev_iommu_fwspec_get(dev);
> +		if (fwspec && iort_pci_rc_supports_ats(node))
>  			fwspec->flags |= IOMMU_FWSPEC_PCI_RC_ATS;
>  	} else {
>  		int i = 0;
Thanks a lot for the fix! I added it to patch 1/14.

Regards,

	Joerg