[RFC PATCH 1/1] x86/VMBus: DMA transfer with encrypted memory in Coco VM

[Tianyu Lan <ltykernel@gmail.com>]

In CoCo VMs, system memory is encrypted by default.
Device drivers typically rely on the DMA core's
SWIOTLB as a bounce buffer for DMA operations, providing
decrypted memory that can be shared between the guest and
host.

For PCI devices with T-Disp support and Confidential
VMBus devices (https://lkml.org/lkml/2026/7/27/1733) can
perform DMA transfers directly with private/encrypted
memory in a CoCo VM.

To support DMA transfer with encrypted memory, Hyper-V
DMA ops are introduced and bypass some API which may
use swiotlb as bounce buffer.

The DMA ops used is global data structure(see get_arch_
dma_ops() and get_dma_ops() for details). There is no
need to set up for each device individually.

Signed-off-by: Tianyu Lan <tiala@microsoft.com>
---
 drivers/hv/Kconfig      |   5 ++
 drivers/hv/Makefile     |   1 +
 drivers/hv/hv_dma_ops.c | 140 ++++++++++++++++++++++++++++++++++++++++
 drivers/hv/vmbus_drv.c  |  14 +++-
 4 files changed, 159 insertions(+), 1 deletion(-)
 create mode 100644 drivers/hv/hv_dma_ops.c

diff --git a/drivers/hv/Kconfig b/drivers/hv/Kconfig
index 2d0b3fcb0ff8..4b50a631fd57 100644
--- a/drivers/hv/Kconfig
+++ b/drivers/hv/Kconfig
@@ -7,6 +7,7 @@ config HYPERV
 	depends on (X86 && X86_LOCAL_APIC && HYPERVISOR_GUEST) \
 		|| (ARM64 && !CPU_BIG_ENDIAN)
 	select PARAVIRT
+	select HYPERV_DMA_OPS
 	select X86_HV_CALLBACK_VECTOR if X86
 	select OF_EARLY_FLATTREE if OF
 	select IRQ_MSI_LIB if X86
@@ -14,6 +15,10 @@ config HYPERV
 	  Select this option to run Linux as a Hyper-V client operating
 	  system.
 
+config HYPERV_DMA_OPS
+	bool "Enable Microsoft Hyper-V DMA ops support"
+	depends on ARCH_HAS_DMA_OPS
+
 config HYPERV_VTL_MODE
 	bool "Enable Linux to boot in VTL context"
 	depends on (X86_64 && HAVE_STATIC_CALL) || ARM64
diff --git a/drivers/hv/Makefile b/drivers/hv/Makefile
index 888a748cc7cb..b4d23c88b9eb 100644
--- a/drivers/hv/Makefile
+++ b/drivers/hv/Makefile
@@ -2,6 +2,7 @@
 obj-$(CONFIG_HYPERV_VMBUS)	+= hv_vmbus.o
 obj-$(CONFIG_HYPERV_UTILS)	+= hv_utils.o
 obj-$(CONFIG_HYPERV_BALLOON)	+= hv_balloon.o
+obj-$(CONFIG_HYPERV_DMA_OPS)	+= hv_dma_ops.o
 obj-$(CONFIG_MSHV_ROOT)		+= mshv_root.o
 obj-$(CONFIG_MSHV_VTL)          += mshv_vtl.o
 
diff --git a/drivers/hv/hv_dma_ops.c b/drivers/hv/hv_dma_ops.c
new file mode 100644
index 000000000000..ef53a42c0450
--- /dev/null
+++ b/drivers/hv/hv_dma_ops.c
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (c) 2026, Microsoft Corporation.
+ *
+ */
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/hyperv.h>
+#include <linux/smpboot.h>
+
+#include <linux/kernel.h>
+#include <linux/dma-map-ops.h>
+#include <linux/export.h>
+#include <asm/mshyperv.h>
+#include "hyperv_vmbus.h"
+#include "../../kernel/dma/direct.h"
+
+extern const struct dma_map_ops *dma_ops;
+
+extern bool is_vmbus_dev(struct device *dev);
+
+static bool hyperv_private_memory_dma(struct device *dev)
+{
+	struct hv_device *hv_dev = device_to_hv_device(dev);
+
+	if (is_vmbus_dev(dev) && hv_dev && hv_dev->channel
+	    && hv_dev->channel->co_external_memory)
+		return true;
+
+	/* Todo: Check T-Disp capability of PCI device here */
+
+	return false;
+}
+
+static int hyperv_dma_map_sg(struct device *dev, struct scatterlist *sgl,
+		int nelems, enum dma_data_direction dir,
+		unsigned long attrs)
+{
+	struct scatterlist *sg;
+	dma_addr_t dma_addr;
+	int i;
+
+	if (hyperv_private_memory_dma(dev)) {
+		for_each_sg(sgl, sg, nelems, i) {
+			dma_addr = __phys_to_dma(dev, sg_phys(sg));
+			sg_dma_address(sg) = dma_addr;
+			sg_dma_len(sg) = sg->length;
+		}
+
+		return nelems;
+	} else {
+		return dma_direct_map_sg(dev, sgl, nelems, dir, attrs);
+	}
+}
+
+static void hyperv_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
+		int nelems, enum dma_data_direction dir, unsigned long attrs)
+{
+	if (!hyperv_private_memory_dma(dev))
+		dma_direct_unmap_sg(dev, sgl, nelems, dir, attrs);
+}
+
+static int hyperv_dma_supported(struct device *dev, u64 mask)
+{
+	dev->coherent_dma_mask = mask;
+	return 1;
+}
+
+static size_t hyperv_dma_max_mapping_size(struct device *dev)
+{
+	if (hyperv_private_memory_dma(dev))
+		return SIZE_MAX;
+	else
+		return swiotlb_max_mapping_size(dev);
+}
+
+/* allocate and map a coherent mapping */
+static void *
+hyperv_dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
+		    gfp_t flag, unsigned long attrs)
+{
+	phys_addr_t phys;
+	void *ret;
+
+	if (!hyperv_private_memory_dma(dev))
+		return dma_alloc_coherent(dev, size, dma_handle, flag);
+
+	size = ALIGN(size, PAGE_SIZE);
+	ret = (void *)__get_free_pages(flag, get_order(size));
+	if (!ret)
+		return ret;
+	phys = virt_to_phys(ret);
+
+	if (hyperv_private_memory_dma(dev))
+		*dma_handle = dma_addr_encrypted(__phys_to_dma(dev, phys));
+	else
+		*dma_handle = phys_to_dma_unencrypted(dev, phys);
+
+	memset(ret, 0, size);
+	return ret;
+}
+
+/* free a coherent mapping */
+static void
+hyperv_dma_free_coherent(struct device *dev, size_t size, void *vaddr,
+		   dma_addr_t dma_addr, unsigned long attrs)
+{
+	if (hyperv_private_memory_dma(dev))
+		dmam_free_coherent(dev, size, vaddr, dma_addr);
+	else
+		free_pages((unsigned long)vaddr, get_order(size));
+}
+
+static dma_addr_t hyperv_dma_map_phys(struct device *dev, phys_addr_t phys,
+		size_t size, enum dma_data_direction dir,
+		unsigned long attrs)
+{
+	if (hyperv_private_memory_dma(dev))
+		return __phys_to_dma(dev, phys);
+	else
+		return dma_direct_map_phys(dev, phys, size, dir, attrs, true);
+}
+
+static void hyperv_dma_unmap_phys(struct device *dev, dma_addr_t dma_handle,
+		size_t size, enum dma_data_direction dir, unsigned long attrs)
+{
+	if (!hyperv_private_memory_dma(dev))
+		dma_direct_unmap_phys(dev, dma_handle, size, dir, attrs, true);
+}
+
+const struct dma_map_ops hyperv_dma_ops = {
+	.alloc			= hyperv_dma_alloc_coherent,
+	.free			= hyperv_dma_free_coherent,
+	.map_phys               = hyperv_dma_map_phys,
+	.unmap_phys             = hyperv_dma_unmap_phys,
+	.map_sg                 = hyperv_dma_map_sg,
+	.unmap_sg               = hyperv_dma_unmap_sg,
+	.dma_supported          = hyperv_dma_supported,
+	.max_mapping_size	= hyperv_dma_max_mapping_size,
+};
diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c
index d28ff45d4cfd..d1ce5c107993 100644
--- a/drivers/hv/vmbus_drv.c
+++ b/drivers/hv/vmbus_drv.c
@@ -40,6 +40,10 @@
 #include <clocksource/hyperv_timer.h>
 #include <asm/mshyperv.h>
 #include "hyperv_vmbus.h"
+#include "../../kernel/dma/direct.h"
+
+extern const struct dma_map_ops *dma_ops;
+extern const struct dma_map_ops hyperv_dma_ops;
 
 struct vmbus_dynid {
 	struct list_head node;
@@ -1468,6 +1472,11 @@ static int vmbus_alloc_synic_and_connect(void)
 	return -ENOMEM;
 }
 
+bool is_vmbus_dev(struct device *dev)
+{
+	return dev->bus == &hv_bus;
+}
+
 /*
  * vmbus_bus_init -Main vmbus driver initialization routine.
  *
@@ -1523,8 +1532,11 @@ static int vmbus_bus_init(void)
 	 * doing that on each VP while initializing SynIC's wastes time.
 	 */
 	is_confidential = ms_hyperv.confidential_vmbus_available;
-	if (is_confidential)
+	if (is_confidential) {
+		dma_ops = &hyperv_dma_ops;
 		pr_info("Establishing connection to the confidential VMBus\n");
+	}
+
 	hv_para_set_sint_proxy(!is_confidential);
 	ret = vmbus_alloc_synic_and_connect();
 	if (ret)
-- 
2.53.0