Hi Vladimir,
Thanks for reviewing.
Hi Christoph Hellwig,
I just want to know if kernel dma mapping/direct is focused on
platforms with MMU.
leave arch code to handle dma coherent memory management themself for
no-MMU platform.
so, you just return error code in kernel/dma/mapping.c,direct.c
without CONFIG_MMU defined ?
which means dma-direct mapping doesn't support !CONFIG_MMU is not a
bug, but design as it's.
or, just return error code currently, will add dma direct mapping
support for !CONFIG_MMU in the
future?
As Vladimir Murzin's suggestion has changes in kernel code, I need
your input to get
the design goal about dma-direct mapping, thanks.
On Tue, Jun 9, 2020 at 10:07 PM Vladimir Murzin <vladimir.murzin@xxxxxxx> wrote:
>
> On 6/8/20 9:30 AM, dillon.minfei@xxxxxxxxx wrote:
> > From: dillon min <dillon.minfei@xxxxxxxxx>
> >
> > Currently, we use dma direct to request coherent memory for driver on armv7m
> > platform if 'cacheid' is zero, but dma_direct_can_mmap() is return false,
> > dma_direct_mmap() return -ENXIO for CONFIG_MMU undefined platform.
> >
> > so we have to back to use 'arm_nommu_dma_ops', add use_reserved_mem() to check
> > if device support global or device corherent memory. if yes, then call
> > set_dma_ops()
> >
> > Signed-off-by: dillon min <dillon.minfei@xxxxxxxxx>
> > ---
> > arch/arm/mm/dma-mapping-nommu.c | 28 +++++++++++++++++++++++++++-
> > 1 file changed, 27 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/mm/dma-mapping-nommu.c b/arch/arm/mm/dma-mapping-nommu.c
> > index 287ef898a55e..e1c213fec152 100644
> > --- a/arch/arm/mm/dma-mapping-nommu.c
> > +++ b/arch/arm/mm/dma-mapping-nommu.c
> > @@ -14,6 +14,7 @@
> > #include <asm/cacheflush.h>
> > #include <asm/outercache.h>
> > #include <asm/cp15.h>
> > +#include <linux/of.h>
> >
> > #include "dma.h"
> >
> > @@ -188,6 +189,31 @@ const struct dma_map_ops arm_nommu_dma_ops = {
> > };
> > EXPORT_SYMBOL(arm_nommu_dma_ops);
> >
> > +static bool use_reserved_mem(struct device *dev)
> > +{
> > + struct device_node *np;
> > +
> > + np = of_find_node_by_path("/reserved-memory/linux,dma");
> > +
> > + if (np &&
> > + of_device_is_compatible(np, "shared-dma-pool") &&
> > + of_property_read_bool(np, "no-map") &&
> > + of_property_read_bool(np, "linux,dma-default")) {
> > + /* has global corherent mem support */
> > + of_node_put(np);
> > + return true;
> > + }
> > +
> > + np = of_parse_phandle(dev->of_node, "memory-region", 0);
> > + if (np) {
> > + /* has dev corherent mem support */
> > + of_node_put(np);
> > + return true;
> > + }
> > +
> > + return false;
> > +}
> > +
> > void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
> > const struct iommu_ops *iommu, bool coherent)
> > {
> > @@ -206,6 +232,6 @@ void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
> > dev->archdata.dma_coherent = (get_cr() & CR_M) ? coherent : true;
> > }
> >
> > - if (!dev->archdata.dma_coherent)
> > + if (!dev->archdata.dma_coherent || use_reserved_mem(dev))
> > set_dma_ops(dev, &arm_nommu_dma_ops);
> > }
> >
>
> Sorry I have to NAK this hack :(
>
> Digging git history reveled 79964a1c2972 ("ARM: 8633/1: nommu: allow mmap when !CONFIG_MMU")
> which make me wonder if diff below does the trick for you
>
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 8f4bbda..8623b9e 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -456,14 +456,14 @@ int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma,
> #else /* CONFIG_MMU */
> bool dma_direct_can_mmap(struct device *dev)
> {
> - return false;
> + return true;
> }
>
> int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma,
> void *cpu_addr, dma_addr_t dma_addr, size_t size,
> unsigned long attrs)
> {
> - return -ENXIO;
> + return vm_iomap_memory(vma, vma->vm_start, (vma->vm_end - vma->vm_start));;
> }
> #endif /* CONFIG_MMU */
Yes, this is a quite nice way to support !CONFIG_MMU without cache on
platforms. I will try your suggestion.
thanks
>
> Cheers
> Vladimir
