Re: [PATCH 1/1] arm64: realm: Use aliased addresses for device DMA to shared buffers

From: Suzuki K Poulose
Date: Wed Feb 12 2025 - 13:48:54 EST

Hi Robin

On 12/02/2025 18:18, Robin Murphy wrote:
On 2025-02-12 5:14 pm, Suzuki K Poulose wrote:
When a device performs DMA to a shared buffer using physical addresses,
(without Stage1 translation), the device must use the "{I}PA address" with the
top bit set in Realm. This is to make sure that a trusted device will be able
to write to shared buffers as well as the protected buffers. Thus, a Realm must
always program the full address including the "protection" bit, like AMD SME
encryption bits.

Add the support for this by providing arm64 version of the phys_to_dma(). We
cannot use the __sme_mask as it assumes the "encryption" always "sets a bit",
which is the opposite for CCA. i.e., "set a bit" for "decrypted" address. So,
move the common code that can be reused by all - i.e., add __phys_to_dma() and
__dma_to_phys() - and do the arch specific processing.

Please note that the VMM needs to similarly make sure that the SMMU Stage2 in
the Non-secure world is setup accordingly to map IPA at the unprotected alias.

Cc: Will Deacon <will@xxxxxxxxxx>
Cc: Jean-Philippe Brucker <jean-philippe@xxxxxxxxxx>
Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
Cc: Robin Murphy <robin.murphy@xxxxxxx>
Cc: Steven Price <steven.price@xxxxxxx>
Cc: Christoph Hellwig <hch@xxxxxx>
Cc: Tom Lendacky <thomas.lendacky@xxxxxxx>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
---
  arch/arm64/Kconfig                  |  1 +
  arch/arm64/include/asm/dma-direct.h | 38 +++++++++++++++++++++++++++++
  include/linux/dma-direct.h          | 35 +++++++++++++++++---------
  3 files changed, 62 insertions(+), 12 deletions(-)
  create mode 100644 arch/arm64/include/asm/dma-direct.h

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index fcdd0ed3eca8..7befe04106de 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -41,6 +41,7 @@ config ARM64
      select ARCH_HAS_NMI_SAFE_THIS_CPU_OPS
      select ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE
      select ARCH_HAS_NONLEAF_PMD_YOUNG if ARM64_HAFT
+    select ARCH_HAS_PHYS_TO_DMA
      select ARCH_HAS_PTE_DEVMAP
      select ARCH_HAS_PTE_SPECIAL
      select ARCH_HAS_HW_PTE_YOUNG
diff --git a/arch/arm64/include/asm/dma-direct.h b/arch/arm64/include/ asm/dma-direct.h
new file mode 100644
index 000000000000..37c3270542b8
--- /dev/null
+++ b/arch/arm64/include/asm/dma-direct.h
@@ -0,0 +1,38 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+#ifndef __ASM_DMA_DIRECT_H
+#define __ASM_DMA_DIRECT_H
+
+#include <asm/pgtable-prot.h>
+
+static inline unsigned long addr_to_shared(unsigned long addr)
+{
+    if (is_realm_world())
+        addr |= prot_ns_shared;
+    return addr;
+}
+
+static inline unsigned long addr_to_private(unsigned long addr)
+{
+    if (is_realm_world())
+        addr &= prot_ns_shared - 1;
+    return addr;
+}
+
+static inline dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
+{
+    return __phys_to_dma(dev, paddr);
+}
+
+static inline dma_addr_t phys_to_dma_unencrypted(struct device *dev,
+                         phys_addr_t paddr)
+{
+    return addr_to_shared(__phys_to_dma(dev, paddr));
+}
+#define phys_to_dma_unencrypted phys_to_dma_unencrypted
+
+static inline phys_addr_t dma_to_phys(struct device *dev, dma_addr_t dma_addr)
+{
+    return addr_to_private(__dma_to_phys(dev, dma_addr));
+}
+
+#endif    /* __ASM_DMA_DIRECT_H */
diff --git a/include/linux/dma-direct.h b/include/linux/dma-direct.h
index d7e30d4f7503..3e9bf6ca640e 100644
--- a/include/linux/dma-direct.h
+++ b/include/linux/dma-direct.h
@@ -72,18 +72,36 @@ static inline dma_addr_t dma_range_map_max(const struct bus_dma_region *map)
      return ret;
  }
+static inline dma_addr_t __phys_to_dma(struct device *dev, phys_addr_t paddr)
+{
+    if (dev->dma_range_map)
+        return translate_phys_to_dma(dev, paddr);
+    return paddr;
+}
+
+static inline phys_addr_t __dma_to_phys(struct device *dev, dma_addr_t dma_addr)
+{
+    phys_addr_t paddr;
+
+    if (dev->dma_range_map)
+        paddr = translate_dma_to_phys(dev, dma_addr);
+    else
+        paddr = dma_addr;
+
+    return paddr;
+}
+
  #ifdef CONFIG_ARCH_HAS_PHYS_TO_DMA
  #include <asm/dma-direct.h>
  #ifndef phys_to_dma_unencrypted
  #define phys_to_dma_unencrypted        phys_to_dma
  #endif
  #else
+
  static inline dma_addr_t phys_to_dma_unencrypted(struct device *dev,
          phys_addr_t paddr)
  {
-    if (dev->dma_range_map)
-        return translate_phys_to_dma(dev, paddr);
-    return paddr;
+    return __phys_to_dma(dev, paddr);
  }
  /*
@@ -94,19 +112,12 @@ static inline dma_addr_t phys_to_dma_unencrypted(struct device *dev,
   */
  static inline dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
  {
-    return __sme_set(phys_to_dma_unencrypted(dev, paddr));
+    return __sme_set(__phys_to_dma(dev, paddr));

Hmm, it really feels like we should generalise __sme_{set,clr} at this level, rather than drag in the entire ARCH_HAS_PHYS_TO_DMA override for the purposes of setting/clearing an address bit just because the "generic" mechanism for doing that is unashamedly AMD-specific.

I was planning to convert that to use ARCH_HAS_PHYS_TO_DMA ;-) too, in a
separate patch, to get rid of this "one approach" works for all hook.


Maybe something like:

#define dma_shared(x)  sme_clr(x)
#define dma_private(x) sme_set(x)

for x86, then go from there?

But this looks sensible too, can switch to this approach.
Thanks for the feedback.

Suzuki



Thanks,
Robin.

  }
  static inline phys_addr_t dma_to_phys(struct device *dev, dma_addr_t dma_addr)
  {
-    phys_addr_t paddr;
-
-    if (dev->dma_range_map)
-        paddr = translate_dma_to_phys(dev, dma_addr);
-    else
-        paddr = dma_addr;
-
-    return __sme_clr(paddr);
+    return __sme_clr(__dma_to_phys(dev, dma_addr));
  }
  #endif /* !CONFIG_ARCH_HAS_PHYS_TO_DMA */