| Message ID | 1382031814-8782-8-git-send-email-stefano.stabellini@eu.citrix.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Russell, are you OK with this patch? On Thu, 17 Oct 2013, Stefano Stabellini wrote: > We can't simply override arm_dma_ops with xen_dma_ops because devices > are allowed to have their own dma_ops and they take precedence over > arm_dma_ops. When running on Xen as initial domain, we always want > xen_dma_ops to be the one in use. > > Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> > Suggested-by: Catalin Marinas <catalin.marinas@arm.com> > CC: will.deacon@arm.com > CC: linux@arm.linux.org.uk > > Changes in v7: > - return xen_dma_ops only if we are the initial domain; > - rename __get_dma_ops to __generic_dma_ops. > --- > arch/arm/include/asm/dma-mapping.h | 13 ++++++++++++- > 1 files changed, 12 insertions(+), 1 deletions(-) > > diff --git a/arch/arm/include/asm/dma-mapping.h b/arch/arm/include/asm/dma-mapping.h > index 01b5a3d..f5945d4 100644 > --- a/arch/arm/include/asm/dma-mapping.h > +++ b/arch/arm/include/asm/dma-mapping.h > @@ -12,17 +12,28 @@ > #include <asm/memory.h> > #include <asm/cacheflush.h> > > +#include <xen/xen.h> > +#include <asm/xen/hypervisor.h> > + > #define DMA_ERROR_CODE (~0) > extern struct dma_map_ops arm_dma_ops; > extern struct dma_map_ops arm_coherent_dma_ops; > > -static inline struct dma_map_ops *get_dma_ops(struct device *dev) > +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > { > if (dev && dev->archdata.dma_ops) > return dev->archdata.dma_ops; > return &arm_dma_ops; > } > > +static inline struct dma_map_ops *get_dma_ops(struct device *dev) > +{ > + if (xen_initial_domain()) > + return xen_dma_ops; > + else > + return __generic_dma_ops(dev); > +} > + > static inline void set_dma_ops(struct device *dev, struct dma_map_ops *ops) > { > BUG_ON(!dev); > -- > 1.7.2.5 >
On Fri, Oct 18, 2013 at 01:22:18PM +0100, Stefano Stabellini wrote: > Russell, > are you OK with this patch? Only concern is why the change is soo large. > > -static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > { > > if (dev && dev->archdata.dma_ops) > > return dev->archdata.dma_ops; > > return &arm_dma_ops; > > } > > > > +static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > +{ > > + if (xen_initial_domain()) > > + return xen_dma_ops; > > + else > > + return __generic_dma_ops(dev); > > +} > > + What's wrong with: static inline struct dma_map_ops *get_dma_ops(struct device *dev) { + if (xen_initial_domain()) + return xen_dma_ops; if (dev && dev->archdata.dma_ops) return dev->archdata.dma_ops; return &arm_dma_ops; }
On Fri, 18 Oct 2013, Russell King - ARM Linux wrote: > On Fri, Oct 18, 2013 at 01:22:18PM +0100, Stefano Stabellini wrote: > > Russell, > > are you OK with this patch? > > Only concern is why the change is soo large. > > > > -static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > > { > > > if (dev && dev->archdata.dma_ops) > > > return dev->archdata.dma_ops; > > > return &arm_dma_ops; > > > } > > > > > > +static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > +{ > > > + if (xen_initial_domain()) > > > + return xen_dma_ops; > > > + else > > > + return __generic_dma_ops(dev); > > > +} > > > + > > What's wrong with: > > static inline struct dma_map_ops *get_dma_ops(struct device *dev) > { > + if (xen_initial_domain()) > + return xen_dma_ops; > if (dev && dev->archdata.dma_ops) > return dev->archdata.dma_ops; > return &arm_dma_ops; > } xen_dma_ops functions might have to call back the native implementation. With the above there is no way for a xen_dma_ops function to get to archdata.dma_ops or arm_dma_ops. This is the reason why this patch introduces __generic_dma_ops.
On Fri, Oct 18, 2013 at 04:46:00PM +0100, Stefano Stabellini wrote: > On Fri, 18 Oct 2013, Russell King - ARM Linux wrote: > > On Fri, Oct 18, 2013 at 01:22:18PM +0100, Stefano Stabellini wrote: > > > Russell, > > > are you OK with this patch? > > > > Only concern is why the change is soo large. > > > > > > -static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > > +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > > > { > > > > if (dev && dev->archdata.dma_ops) > > > > return dev->archdata.dma_ops; > > > > return &arm_dma_ops; > > > > } > > > > > > > > +static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > > +{ > > > > + if (xen_initial_domain()) > > > > + return xen_dma_ops; > > > > + else > > > > + return __generic_dma_ops(dev); > > > > +} > > > > + > > > > What's wrong with: > > > > static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > { > > + if (xen_initial_domain()) > > + return xen_dma_ops; > > if (dev && dev->archdata.dma_ops) > > return dev->archdata.dma_ops; > > return &arm_dma_ops; > > } > > > xen_dma_ops functions might have to call back the native implementation. > With the above there is no way for a xen_dma_ops function to get to > archdata.dma_ops or arm_dma_ops. > This is the reason why this patch introduces __generic_dma_ops. Ah, that's fine then, but having that described in the commit message would be useful. Acked-by: Russell King <rmk+kernel@arm.linux.org.uk> Thanks.
On Fri, 18 Oct 2013, Russell King - ARM Linux wrote: > On Fri, Oct 18, 2013 at 04:46:00PM +0100, Stefano Stabellini wrote: > > On Fri, 18 Oct 2013, Russell King - ARM Linux wrote: > > > On Fri, Oct 18, 2013 at 01:22:18PM +0100, Stefano Stabellini wrote: > > > > Russell, > > > > are you OK with this patch? > > > > > > Only concern is why the change is soo large. > > > > > > > > -static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > > > +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > > > > { > > > > > if (dev && dev->archdata.dma_ops) > > > > > return dev->archdata.dma_ops; > > > > > return &arm_dma_ops; > > > > > } > > > > > > > > > > +static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > > > +{ > > > > > + if (xen_initial_domain()) > > > > > + return xen_dma_ops; > > > > > + else > > > > > + return __generic_dma_ops(dev); > > > > > +} > > > > > + > > > > > > What's wrong with: > > > > > > static inline struct dma_map_ops *get_dma_ops(struct device *dev) > > > { > > > + if (xen_initial_domain()) > > > + return xen_dma_ops; > > > if (dev && dev->archdata.dma_ops) > > > return dev->archdata.dma_ops; > > > return &arm_dma_ops; > > > } > > > > > > xen_dma_ops functions might have to call back the native implementation. > > With the above there is no way for a xen_dma_ops function to get to > > archdata.dma_ops or arm_dma_ops. > > This is the reason why this patch introduces __generic_dma_ops. > > Ah, that's fine then, but having that described in the commit message > would be useful. > > Acked-by: Russell King <rmk+kernel@arm.linux.org.uk> > > Thanks. Thank you. I'll add a note to the commit message.
diff --git a/arch/arm/include/asm/dma-mapping.h b/arch/arm/include/asm/dma-mapping.h index 01b5a3d..f5945d4 100644 --- a/arch/arm/include/asm/dma-mapping.h +++ b/arch/arm/include/asm/dma-mapping.h @@ -12,17 +12,28 @@ #include <asm/memory.h> #include <asm/cacheflush.h> +#include <xen/xen.h> +#include <asm/xen/hypervisor.h> + #define DMA_ERROR_CODE (~0) extern struct dma_map_ops arm_dma_ops; extern struct dma_map_ops arm_coherent_dma_ops; -static inline struct dma_map_ops *get_dma_ops(struct device *dev) +static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) { if (dev && dev->archdata.dma_ops) return dev->archdata.dma_ops; return &arm_dma_ops; } +static inline struct dma_map_ops *get_dma_ops(struct device *dev) +{ + if (xen_initial_domain()) + return xen_dma_ops; + else + return __generic_dma_ops(dev); +} + static inline void set_dma_ops(struct device *dev, struct dma_map_ops *ops) { BUG_ON(!dev);
We can't simply override arm_dma_ops with xen_dma_ops because devices are allowed to have their own dma_ops and they take precedence over arm_dma_ops. When running on Xen as initial domain, we always want xen_dma_ops to be the one in use. Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com> Suggested-by: Catalin Marinas <catalin.marinas@arm.com> CC: will.deacon@arm.com CC: linux@arm.linux.org.uk Changes in v7: - return xen_dma_ops only if we are the initial domain; - rename __get_dma_ops to __generic_dma_ops. --- arch/arm/include/asm/dma-mapping.h | 13 ++++++++++++- 1 files changed, 12 insertions(+), 1 deletions(-)