[PATCH kernel] powerpc/iommu: Report the correct most efficient DMA mask for PCI devices

According to dma-api.rst, the dma_get_required_mask() helper should return
"the mask that the platform requires to operate efficiently". Which in
the case of PPC64 means the bypass mask and not a mask from an IOMMU table
which is shorter and slower to use due to map/unmap operations (especially
expensive on "pseries").

However the existing implementation ignores the possibility of bypassing
and returns the IOMMU table mask on the pseries platform which makes some
drivers (mpt3sas is one example) choose 32bit DMA even though bypass is
supported. The powernv platform sort of handles it by having a bigger
default window with a mask >=40 but it only works as drivers choose
63/64bit if the required mask is >32 which is rather pointless.

This reintroduces the bypass capability check to let drivers make
a better choice of the DMA mask.

Fixes: f1565c24b596 ("powerpc: use the generic dma_ops_bypass mode")
Signed-off-by: Alexey Kardashevskiy <redacted>
---
 arch/powerpc/kernel/dma-iommu.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/arch/powerpc/kernel/dma-iommu.c b/arch/powerpc/kernel/dma-iommu.c
index 111249fd619d..d646077bcbcf 100644
--- a/arch/powerpc/kernel/dma-iommu.c
+++ b/arch/powerpc/kernel/dma-iommu.c

@@ -184,6 +184,14 @@ u64 dma_iommu_get_required_mask(struct device *dev)
 	struct iommu_table *tbl = get_iommu_table_base(dev);
 	u64 mask;
 
+	if (dev_is_pci(dev)) {
+		u64 bypass_mask = dma_direct_get_required_mask(dev);
+
+		if (dma_iommu_dma_supported(dev, bypass_mask)) {
+			dev_info(dev, "%s: returning bypass mask 0x%llx\n", __func__, bypass_mask);
+			return bypass_mask;
+		}
+	}
 	if (!tbl)
 		return 0;
 

-- 
2.30.2

On Thu, Sep 30, 2021 at 01:44:54PM +1000, Alexey Kardashevskiy wrote:

and returns the IOMMU table mask on the pseries platform which makes some
drivers (mpt3sas is one example) choose 32bit DMA even though bypass is
supported. The powernv platform sort of handles it by having a bigger
default window with a mask >=40 but it only works as drivers choose
63/64bit if the required mask is >32 which is rather pointless.

This reintroduces the bypass capability check to let drivers make
a better choice of the DMA mask.

Fixes: f1565c24b596 ("powerpc: use the generic dma_ops_bypass mode")
Signed-off-by: Alexey Kardashevskiy <redacted>

Looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>

On Thu, 30 Sep 2021 13:44:54 +1000, Alexey Kardashevskiy wrote:

According to dma-api.rst, the dma_get_required_mask() helper should return
"the mask that the platform requires to operate efficiently". Which in
the case of PPC64 means the bypass mask and not a mask from an IOMMU table
which is shorter and slower to use due to map/unmap operations (especially
expensive on "pseries").

However the existing implementation ignores the possibility of bypassing
and returns the IOMMU table mask on the pseries platform which makes some
drivers (mpt3sas is one example) choose 32bit DMA even though bypass is
supported. The powernv platform sort of handles it by having a bigger
default window with a mask >=40 but it only works as drivers choose
63/64bit if the required mask is >32 which is rather pointless.

[...]

Applied to powerpc/fixes.

[1/1] powerpc/iommu: Report the correct most efficient DMA mask for PCI devices
      https://git.kernel.org/powerpc/c/23c216b335d1fbd716076e8263b54a714ea3cf0e

cheers