| Message ID | 516b33118d489e56499ff8c64c019709b744110c.1597931876.git.robin.murphy@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Convert arch/arm to use iommu-dma | expand |
20.08.2020 18:08, Robin Murphy пишет: > Now that arch/arm is wired up for default domains and iommu-dma, > implement the corresponding driver-side support for DMA domains. > > Signed-off-by: Robin Murphy <robin.murphy@arm.com> > --- > drivers/iommu/tegra-gart.c | 17 ++++++++++++----- > 1 file changed, 12 insertions(+), 5 deletions(-) > > diff --git a/drivers/iommu/tegra-gart.c b/drivers/iommu/tegra-gart.c > index fac720273889..e081387080f6 100644 > --- a/drivers/iommu/tegra-gart.c > +++ b/drivers/iommu/tegra-gart.c > @@ -9,6 +9,7 @@ > > #define dev_fmt(fmt) "gart: " fmt > > +#include <linux/dma-iommu.h> > #include <linux/io.h> > #include <linux/iommu.h> > #include <linux/moduleparam.h> > @@ -145,16 +146,22 @@ static struct iommu_domain *gart_iommu_domain_alloc(unsigned type) > { > struct iommu_domain *domain; Hello, Robin! Tegra20 GART isn't a real IOMMU, but a small relocation aperture. We would only want to use it for a temporal mappings (managed by GPU driver) for the time while GPU hardware is busy and working with a sparse DMA buffers, the driver will take care of unmapping the sparse buffers once GPU work is finished [1]. In a case of contiguous DMA buffers, we want to bypass the IOMMU and use buffer's phys address because GART aperture is small and all buffers simply can't fit into GART for a complex GPU operations that involve multiple buffers [2][3]. The upstream GPU driver still doesn't support GART, but eventually it needs to be changed. [1] https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/gart.c#L489 [2] https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/gart.c#L542 [3] https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/uapi/patching.c#L90 > - if (type != IOMMU_DOMAIN_UNMANAGED) > + if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA) > return NULL; Will a returned NULL tell to IOMMU core that implicit domain shouldn't be used? Is it possible to leave this driver as-is?
On 2020-08-20 21:16, Dmitry Osipenko wrote: > 20.08.2020 18:08, Robin Murphy пишет: >> Now that arch/arm is wired up for default domains and iommu-dma, >> implement the corresponding driver-side support for DMA domains. >> >> Signed-off-by: Robin Murphy <robin.murphy@arm.com> >> --- >> drivers/iommu/tegra-gart.c | 17 ++++++++++++----- >> 1 file changed, 12 insertions(+), 5 deletions(-) >> >> diff --git a/drivers/iommu/tegra-gart.c b/drivers/iommu/tegra-gart.c >> index fac720273889..e081387080f6 100644 >> --- a/drivers/iommu/tegra-gart.c >> +++ b/drivers/iommu/tegra-gart.c >> @@ -9,6 +9,7 @@ >> >> #define dev_fmt(fmt) "gart: " fmt >> >> +#include <linux/dma-iommu.h> >> #include <linux/io.h> >> #include <linux/iommu.h> >> #include <linux/moduleparam.h> >> @@ -145,16 +146,22 @@ static struct iommu_domain *gart_iommu_domain_alloc(unsigned type) >> { >> struct iommu_domain *domain; > > Hello, Robin! > > Tegra20 GART isn't a real IOMMU, but a small relocation aperture. We > would only want to use it for a temporal mappings (managed by GPU > driver) for the time while GPU hardware is busy and working with a > sparse DMA buffers, the driver will take care of unmapping the sparse > buffers once GPU work is finished [1]. In a case of contiguous DMA > buffers, we want to bypass the IOMMU and use buffer's phys address > because GART aperture is small and all buffers simply can't fit into > GART for a complex GPU operations that involve multiple buffers [2][3]. > The upstream GPU driver still doesn't support GART, but eventually it > needs to be changed. > > [1] > https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/gart.c#L489 > > [2] > https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/gart.c#L542 > > [3] > https://github.com/grate-driver/linux/blob/master/drivers/gpu/drm/tegra/uapi/patching.c#L90 > >> - if (type != IOMMU_DOMAIN_UNMANAGED) >> + if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA) >> return NULL; > > Will a returned NULL tell to IOMMU core that implicit domain shouldn't > be used? Is it possible to leave this driver as-is? The aim of this patch was just to make the conversion without functional changes wherever possible, i.e. maintain an equivalent to the existing ARM behaviour of allocating its own implicit domains for everything. It doesn't represent any judgement of whether that was ever appropriate for this driver in the first place ;) Hopefully my other reply already covered the degree of control drivers can have with proper default domains, but do shout if anything wasn't clear. Cheers, Robin.
21.08.2020 03:28, Robin Murphy пишет: ... >> Will a returned NULL tell to IOMMU core that implicit domain shouldn't >> be used? Is it possible to leave this driver as-is? > > The aim of this patch was just to make the conversion without functional > changes wherever possible, i.e. maintain an equivalent to the existing > ARM behaviour of allocating its own implicit domains for everything. It > doesn't represent any judgement of whether that was ever appropriate for > this driver in the first place ;) > > Hopefully my other reply already covered the degree of control drivers > can have with proper default domains, but do shout if anything wasn't > clear. Thank you for the detailed comments! I wasn't watching closely all the recent iommu/ changes and yours clarification is very helpful! My current understanding is that the GART driver will need to support the IOMMU_DOMAIN_IDENTITY and set def_domain_type to IOMMU_DOMAIN_IDENTITY for all devices. Meanwhile, today's upstream drivers don't use GART, hence this patch should be okay. Although, it's a bit unlikely that the IOMMU_DOMAIN_DMA type will ever be useful for the GART, and thus, I'm still thinking that will be a bit nicer to keep GART driver as-is for now.
diff --git a/drivers/iommu/tegra-gart.c b/drivers/iommu/tegra-gart.c index fac720273889..e081387080f6 100644 --- a/drivers/iommu/tegra-gart.c +++ b/drivers/iommu/tegra-gart.c @@ -9,6 +9,7 @@ #define dev_fmt(fmt) "gart: " fmt +#include <linux/dma-iommu.h> #include <linux/io.h> #include <linux/iommu.h> #include <linux/moduleparam.h> @@ -145,16 +146,22 @@ static struct iommu_domain *gart_iommu_domain_alloc(unsigned type) { struct iommu_domain *domain; - if (type != IOMMU_DOMAIN_UNMANAGED) + if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA) return NULL; domain = kzalloc(sizeof(*domain), GFP_KERNEL); - if (domain) { - domain->geometry.aperture_start = gart_handle->iovmm_base; - domain->geometry.aperture_end = gart_handle->iovmm_end - 1; - domain->geometry.force_aperture = true; + if (!domain) + return NULL; + + if (type == IOMMU_DOMAIN_DMA && iommu_get_dma_cookie(domain)) { + kfree(domain); + return NULL; } + domain->geometry.aperture_start = gart_handle->iovmm_base; + domain->geometry.aperture_end = gart_handle->iovmm_end - 1; + domain->geometry.force_aperture = true; + return domain; }
Now that arch/arm is wired up for default domains and iommu-dma, implement the corresponding driver-side support for DMA domains. Signed-off-by: Robin Murphy <robin.murphy@arm.com> --- drivers/iommu/tegra-gart.c | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-)