iommu/of: Deconfigure iommu on driver detach
Subsystems (2)
RFC PATCH 1/1 iommu/of: Deconfigure iommu on driver detach
As part of dma_deconfigure, lets deconfigure the iommu too on driver detach, so that we clear the iommu domain and related group.
---
As a part of dma_deconfigure, shouldn't we deconfigure the iommu as well? This should reverse all that we do in of_iommu_configure. So that we call the .remove_device ops for iommu and eventually clear all the iommu domain, related group infrastructure.
I am seeing that the loadable modules get the same iommu configurations after re-loading them, i.e. iommu domain, and iova_domain configurations, as we didn't cleared them.
So should we be clearing these configurations, and therefore do a of_iommu_deconfigure() or sort?
drivers/iommu/of_iommu.c | 11 +++++++++++ drivers/of/device.c | 1 + include/linux/of_iommu.h | 5 +++++ 3 files changed, 17 insertions(+)
drivers/iommu/of_iommu.c +11 -0
@@ -160,6 +160,17 @@ static int of_pci_iommu_init(struct pci_dev *pdev, u16 alias, void *data)return err; } +void of_iommu_deconfigure (struct device *dev) +{ +const struct iommu_ops *ops =NULL ; + +if (dev->iommu_fwspec && dev->iommu_fwspec->ops) + ops = dev->iommu_fwspec->ops; + +if (ops && ops->remove_device && dev->iommu_group) + ops->remove_device (dev); +} +const struct iommu_ops *of_iommu_configure (struct device *dev,struct device_node *master_np) {
drivers/of/device.c +1 -0
@@ -177,6 +177,7 @@ EXPORT_SYMBOL_GPL(of_dma_configure);void of_dma_deconfigure (struct device *dev) {arch_teardown_dma_ops (dev); +of_iommu_deconfigure (dev); }int of_device_register (struct platform_device *pdev)
include/linux/of_iommu.h +5 -0
@@ -15,6 +15,8 @@ extern int of_get_dma_window(struct device_node *dn, const char *prefix,extern const struct iommu_ops *of_iommu_configure (struct device *dev,struct device_node *master_np); +extern void of_iommu_deconfigure (struct device *dev); +#else static inline int of_get_dma_window(struct device_node *dn, const char *prefix,
@@ -30,6 +32,9 @@ static inline const struct iommu_ops *of_iommu_configure(struct device *dev, return NULL; } +static inline void of_iommu_deconfigure (struct device *dev) +{ } +#endif /* CONFIG_OF_IOMMU */ extern struct of_device_id __iommu_of_table;
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation
A message that is not in the archive (sent off-list or not yet received): 20180328102529.13435-1-vivek.gautam-sgV2jX0FEOL9JmXXK+q4OQ@public.gmane.org
Quoted 7 lines · Vivek Gautam
As part of
dma_deconfigure, lets deconfigure the iommu too on driver detach, so that we clear the iommu domain and related group.Signed-off-by Vivek Gautam <[email protected]>---
Quoted 9 lines
As a part of
dma_deconfigure, shouldn't we deconfigure the iommu as well? This should reverse all that we do inof_iommu_configure. So that we call the .remove_device ops for iommu and eventually clear all the iommu domain, related group infrastructure.
I am seeing that the loadable modules get the same iommu configurations after re-loading them, i.e. iommu domain, andiova_domainconfigurations, as we didn't cleared them.
So should we be clearing these configurations, and therefore do aof_iommu_deconfigure()or sort?
Nope. Unbinding a driver does not cause the device to stop existing, nor remove it from its parent bus, which is what .remove_device represents. Device grouping is based on the underlying hardware topology, which isn't going to change at runtime, so there's little point in the software state pretending otherwise. If a module is leaking DMA mappings in the default domain when unloaded, that's a bug in that module (and certainly not specific to the IOMMU layer).
For reference, the only reason we touch .add_device in of_iommu_configure() is because iommu_bus_init() is not quite reliable enough for multiple IOMMU instances serving the platform bus (or similar). It's an ugly workaround which should go away if and when we manage to get rid of the notion of per-bus ops in the API.
Robin.
Quoted 64 lines
drivers/iommu/of_iommu.c | 11 +++++++++++ drivers/of/device.c | 1 + include/linux/of_iommu.h | 5 +++++ 3 files changed, 17 insertions(+)drivers/iommu/of_iommu.c +11
-0@@ -160,6 +160,17 @@ static int of_pci_iommu_init(struct pci_dev *pdev, u16 alias, void *data)return err; } +void of_iommu_deconfigure (struct device *dev) +{ +const struct iommu_ops *ops =NULL ; + +if (dev->iommu_fwspec && dev->iommu_fwspec->ops) + ops = dev->iommu_fwspec->ops; + +if (ops && ops->remove_device && dev->iommu_group) + ops->remove_device (dev); +} +const struct iommu_ops *of_iommu_configure (struct device *dev,struct device_node *master_np) {drivers/of/device.c +1
-0@@ -177,6 +177,7 @@ EXPORT_SYMBOL_GPL(of_dma_configure);void of_dma_deconfigure (struct device *dev) {arch_teardown_dma_ops (dev); +of_iommu_deconfigure (dev); }int of_device_register (struct platform_device *pdev)include/linux/of_iommu.h +5
-0@@ -15,6 +15,8 @@ extern int of_get_dma_window(struct device_node *dn, const char *prefix,extern const struct iommu_ops *of_iommu_configure (struct device *dev,struct device_node *master_np); +extern void of_iommu_deconfigure (struct device *dev); +#else static inline int of_get_dma_window(struct device_node *dn, const char *prefix,@@ -30,6 +32,9 @@ static inline const struct iommu_ops *of_iommu_configure(struct device *dev, return NULL; } +static inline void of_iommu_deconfigure (struct device *dev) +{ } +#endif /* CONFIG_OF_IOMMU */ extern struct of_device_id __iommu_of_table;
A message that is not in the archive (sent off-list or not yet received): 993dcde3-f518-9852-9271-3a2c5a3300a5-5wv7dgnIgG8@public.gmane.org
Hi Robin,
Quoted 18 lines · Robin Murphy
Quoted 7 lines · Vivek Gautam
As part of
dma_deconfigure, lets deconfigure the iommu too on driver detach, so that we clear the iommu domain and related group.Signed-off-by Vivek Gautam <[email protected]>---
Quoted 9 lines
As a part of
dma_deconfigure, shouldn't we deconfigure the iommu as well? This should reverse all that we do inof_iommu_configure. So that we call the .remove_device ops for iommu and eventually clear all the iommu domain, related group infrastructure.
I am seeing that the loadable modules get the same iommu configurations after re-loading them, i.e. iommu domain, andiova_domainconfigurations, as we didn't cleared them.
So should we be clearing these configurations, and therefore do aof_iommu_deconfigure()or sort?
Quoted 1 line
Nope. Unbinding a driver does not cause the device to stop existing, nor remove it from its parent bus, which is what .remove_device represents. Device grouping is based on the underlying hardware topology, which isn't going to change at runtime, so there's little point in the software state pretending otherwise. If a module is leaking DMA mappings in the default domain when unloaded, that's a bug in that module (and certainly not specific to the IOMMU layer).
Thanks for your review, and really sorry for the delayed response.
I can follow what you said. So the device will only be removed (by .remove_device) when there's a BUS_NOTIFY_REMOVED_DEVICE notifier call. We can drop this patch.
Best regards
Vivek
Quoted 58 lines
For reference, the only reason we touch .add_device in
of_iommu_configure()is becauseiommu_bus_init()is not quite reliable enough for multiple IOMMU instances serving the platform bus (or similar). It's an ugly workaround which should go away if and when we manage to get rid of the notion of per-bus ops in the API.Robin.
Quoted 53 lines
drivers/iommu/of_iommu.c| 11 +++++++++++
drivers/of/device.c | 1 +
include/linux/of_iommu.h| 5 +++++3 files changed, 17 insertions(+)drivers/iommu/of_iommu.c +10
-0@@ -160,6 +160,17 @@ static int of_pci_iommu_init(struct pci_dev *pdev, u16 alias, void *data) return err; } +void of_iommu_deconfigure(struct device *dev) +{ + const struct iommu_ops *ops = NULL; + + if (dev->iommu_fwspec && dev->iommu_fwspec->ops) + ops = dev->iommu_fwspec->ops; + + if (ops && ops->remove_device && dev->iommu_group) + ops->remove_device(dev); +} + const struct iommu_ops *of_iommu_configure(struct device *dev, struct device_node *master_np){
drivers/of/device.c +1
-0@@ -177,6 +177,7 @@ EXPORT_SYMBOL_GPL(of_dma_configure); void of_dma_deconfigure(struct device *dev) { arch_teardown_dma_ops(dev); + of_iommu_deconfigure(dev); } int of_device_register(struct platform_device *pdev)include/linux/of_iommu.h +1
-0@@ -15,6 +15,8 @@ extern int of_get_dma_window(struct device_node *dn, const char *prefix, extern const struct iommu_ops *of_iommu_configure(struct device *dev, struct device_node *master_np); +extern void of_iommu_deconfigure(struct device *dev); + #else static inline int of_get_dma_window(struct device_node *dn, const char *prefix,@@ -30,6 +32,9 @@ static inline const struct iommu_ops *of_iommu_configure(struct device *dev,
Quoted 8 lines
include/linux/of_iommu.h static inline const struct iommu_ops *of_iommu_configure(struct device *dev, · line 40
Quoted 8 lines
include/linux/of_iommu.h +2
-0return NULL; } +static inline void of_iommu_deconfigure(struct device *dev) +{ } + #endif /* CONFIG_OF_IOMMU */ extern struct of_device_id __iommu_of_table;
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu