On 2026-02-11 6:00 pm, Michael Kelley wrote:
From: Tianyu Lan <[email protected]> Sent: Tuesday, February 10, 2026 8:21 AM
Hyper-V provides Confidential VMBus to communicate between
device model and device guest driver via encrypted/private
memory in Confidential VM. The device model is in OpenHCL
(https://openvmm.dev/guide/user_guide/openhcl.html) that
plays the paravisor rule.
For a VMBUS device, there are two communication methods to
s/VMBUS/VMBus/
talk with Host/Hypervisor. 1) VMBus Ring buffer 2) dynamic
DMA transition.
I'm not sure what "dynamic DMA transition" is. Maybe just
"DMA transfers"? Also, do the same substitution further
down in this commit message.
The Confidential VMBus Ring buffer has been
upstreamed by Roman Kisel(commit 6802d8af).
It's customary to use 12 character commit IDs, which would be
6802d8af47d1 in this case.
The dynamic DMA transition of VMBus device normally goes
through DMA core and it uses SWIOTLB as bounce buffer in
CVM
"CVM" is Microsoft-speak. The Linux terminology is "a CoCo VM".
to communicate with Host/Hypervisor. The Confidential
VMBus device may use private/encrypted memory to do DMA
and so the device swiotlb(bounce buffer) isn't necessary.
The phrase "isn't necessary" does not capture the real issue
here. Saying "isn't necessary" makes it sound like this patch is
just avoids unnecessary work, so that it is a performance
improvement. But that's not the case.
The real issue is that swiotlb memory is decrypted. So bouncing
through the swiotlb exposes to the host what is supposed to be
confidential data passed on the Confidential VMBus. Disabling
the swiotlb bouncing in this case is a hard requirement to preserve
confidentially.
Yeah, this really isn't a Hyper-V problem. Indeed as things stand,
"swiotlb=force" could potentially break confidentiality for any
environment trying to invent a notion of private DMA, and perhaps we
could throw a big warning about that, but really the answer there is
"Don't run your confidential workload with 'swiotlb=force'. Why would
you even do that? Debug your drivers in a regular VM or bare-metal with
full debug visibility like a normal person..."
The fact is we do not have a proper notion of trusted/private DMA yet,
and this is not the way to add it. The current assumption is very much
that all DMA is untrusted in the CoCo sense, because initially it was
only virtual devices emulated by a hypervisor, thus had to be bounced
through shared memory anyway. AMD SEV with a stage 1 IOMMU in the guest
can allow an assigned physical device to access a suitably-aligned
encrypted buffer directly, but that's still effectively just putting the
buffer into a temporarily shared state for that device, it merely skips
sharing it with the rest of the system. !force_dma_unencrypted() doesn't
mean "we trust this device's DMA", it just means "we don't have to use
explicitly-decrypted pages to accommodate untrusted/shared DMA here",
plus it also serves double-duty for host encryption which doesn't share
the same trust model anyway.
I assumed this would follow the TDISP stuff, but if Hyper-V has an
alternative device-trusting mechanism already then there's no need to
wait. We want some common device property (likely consolidating the
current PCI external-facing port notion of trustedness plus whatever
TDISP wants), with which we can then make proper decisions in all the
right DMA API paths - and if it can end up replacing the horrible
force_dma_unencrypted() as well then all the better! I'd totally
forgotten about the previous discussion that Michael referred to (which
I had to track down[1]), but it looks like all the main points were
already covered there and we were approaching a consensus, so really I
guess someone just needs to give it a go.
Thanks,
Robin.
[1]
https://lore.kernel.org/linux-hyperv/[email protected]/
So I would reword the sentences as something like this:
The Confidential VMBus device can do DMA directly to
private/encrypted memory. Because the swiotlb is decrypted
memory, the DMA transfer must not be bounced through the
swiotlb, so as to preserve confidentiality. This is different from
the default for Linux CoCo VMs, so disable the VMBus device's
use of swiotlb.
To disable device's swiotlb, set device->dma_io_tlb_mem
to NULL in VMBus driver and is_swiotlb_force_bounce()
always returns false.
Suggested-by: Michael Kelley <[email protected]>
Signed-off-by: Tianyu Lan <[email protected]>
---
Change since v1:
Use device.dma_io_tlb_mem to disable device bounce buffer
drivers/hv/vmbus_drv.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c
index a53af6fe81a6..58dab8cc3fcb 100644
--- a/drivers/hv/vmbus_drv.c
+++ b/drivers/hv/vmbus_drv.c
@@ -2133,11 +2133,15 @@ int vmbus_device_register(struct hv_device
*child_device_obj)
child_device_obj->device.dma_mask = &child_device_obj->dma_mask;
dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64));
+ device_initialize(&child_device_obj->device);
+ if (child_device_obj->channel->co_external_memory)
+ child_device_obj->device.dma_io_tlb_mem = NULL;
+
Doing this as part of the VMBus bus driver makes sense. While directly
setting device.dma_io_tlb_mem to NULL should work, it would be better
to add a function to the swiotlb code to do this, and then call that function
here, passing the device as an argument. The need to disable swiotlb on a
device will likely arise in similar contexts (such as TDISP), and it would be
better to have a swiotlb function for that purpose. This use case may be
a bit ahead of the TDISP work, and having a swiotlb function in place will
help ensure that duplicate mechanisms aren't created as everything
comes together.
See my earlier comments in [1] about the key point in the commit message,
and about adding a swiotlb_dev_disable() function to the swiotlb code.
Michael
[1]
https://lore.kernel.org/linux-hyperv/sn6pr02mb4157dae6d8cc6ba11ca87298d4...@sn6pr02mb4157.namprd02.prod.outlook.com/
/*
* Register with the LDM. This will kick off the driver/device
* binding...which will eventually call vmbus_match() and vmbus_probe()
*/
- ret = device_register(&child_device_obj->device);
+ ret = device_add(&child_device_obj->device);
if (ret) {
pr_err("Unable to register child device\n");
put_device(&child_device_obj->device);
--
2.50.1
