Re: [PATCH v2 18/19] PCI: Replace untrusted state with device trust

From: Robin Murphy

Date: Mon Nov 16 2026 - 12:14:35 EST


On 10/11/2026 12:22 pm, Leon Romanovsky wrote:
On Mon, Nov 09, 2026 at 01:53:02PM +0000, Robin Murphy wrote:
On 09/11/2026 7:33 am, Leon Romanovsky wrote:
[...]
diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
index b486b8bbd1fcd..8be2a246d4903 100644
--- a/drivers/iommu/iommu.c
+++ b/drivers/iommu/iommu.c
@@ -1879,7 +1879,7 @@ static int iommu_get_def_domain_type(struct iommu_group *group,
static int iommu_get_default_domain_type(struct iommu_group *group,
int target_type)
{
- struct device *untrusted = NULL;
+ struct device *restricted = NULL;
struct group_device *gdev;
int driver_type = 0;
@@ -1901,14 +1901,15 @@ static int iommu_get_default_domain_type(struct iommu_group *group,
driver_type = iommu_get_def_domain_type(group, gdev->dev,
driver_type);
- if (dev_is_pci(gdev->dev) && to_pci_dev(gdev->dev)->untrusted) {
+ if (dev_is_pci(gdev->dev) &&
+ pci_dev_default_is_restricted(to_pci_dev(gdev->dev))) {

I think this should be the generic device_is_adversarial(gdev->dev) now - in
fact probably all the dev_is_pci()/to_pci_dev() stanzas aren't conceptually
PCI-specific, just an artefact of the current "untrusted" flag being so.

I wanted to do the same, but iommu_get_default_domain_type() may be called
before the PCI device's trust level has been set. At that point, we are
still in "disabled" mode, whereas device_is_adversarial() checks for
"adversarial" mode.

The below is from AI, I asked to present me the path for that:
--------------------------------------------------------------------------------
For a newly discovered PCI device with an available IOMMU and a new IOMMU group:
pci_scan_single_device()
pci_device_add()
device_add()
bus_notify(dev, BUS_NOTIFY_ADD_DEVICE)
blocking_notifier_call_chain()
notifier_call_chain()
iommu_bus_notifier()
iommu_probe_device(dev)
__iommu_probe_device(dev, NULL)
iommu_setup_default_domain(group, 0)
iommu_get_default_domain_type(group, 0)

For a device whose initial policy is adversary, its state during that call is:
dev->trust_policy = DEVICE_TRUST_POLICY_ADVERSARY
dev->trust_level = DEVICE_TRUST_DISABLED
dev->trust_committed_level = DEVICE_TRUST_DISABLED
dev->driver = NULL

device_is_adversarial(dev) = false

The active level changes later, during binding:
really_probe()
device_trust_prepare()
__device_trust_prepare()
WRITE_ONCE(dev->trust_level, DEVICE_TRUST_ADVERSARY)
dev->bus->dma_configure()
call_driver_probe()

OK, that seems less than ideal. Feels like it would be better if bus code set the initial device state prior to device_add, so then it only ever changes later as a result of an explicit attestation/acceptance process. If the initial IOMMU probe still has to poke into bus-specific code to avoid always making a worst-case assumption, doesn't that rather defeat the point of having a generic abstraction?

Even with the slightly dubious argument that this series only intends to shuffle the existing PCI code around rather than fully introduce the generic abstraction that it claims to, if the PCI behaviour is still just based on propagating the internal "untrusted" flag plus the new global override, both of which *are* known at add_device time into the common trust state, then there seems little excuse for not doing that up-front so that the new API *can* just work as expected.


/*
- * No ARM32 using systems will set untrusted, it cannot
- * work.
+ * No ARM32 using systems select a restricted trust
+ * policy; it cannot work.
*/
if (WARN_ON(IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)))
return -1;
- untrusted = gdev->dev;
+ restricted = gdev->dev;
}
}
[...]
@@ -616,6 +608,28 @@ static inline struct pci_dev *to_pci_dev(const struct device *dev)
{
return container_of(dev, struct pci_dev, dev);
}
+
+static inline bool
+pci_dev_default_is_restricted(const struct pci_dev *pdev)
+{
+ enum device_trust_policy policy = READ_ONCE(pdev->dev.trust_policy);
+
+ return policy == DEVICE_TRUST_POLICY_DISABLED ||
+ policy == DEVICE_TRUST_POLICY_ADVERSARY;
+}

Unless I've wildly misunderstood, I'm not sure this is right - if the user
has set DEVICE_TRUST_POLICY_DISABLED to explicitly override the bus-default
behaviour, but that leads to the IOMMU layer special-casing all PCI devices
as adversarial and thus force-*enabling* all their trust-model restrictions,
isn't that pretty much the complete opposite of what was asked for?

We have two levels: policy - that describes the default of all devices
for that bus. level - actual decision on how that device is going to be
operated.

"Disabled" prevents driver probing until the device is authorized,
"Adversary" permits probing and requests the available protections
for operating an untrusted device.

Both need conservative initial IOMMU setup: a disabled device may later
be authorized as adversary, and authorization does not rebuild its IOMMU
domain.

Urgh, I can be absolutely certain I'm not the only one who is going to be confused this way... I see "disabled" and instinctively take it to mean disabling the device_trust *mechanism*, as the logical thing that people will reach for in terms of "I don't care what this new thing is, just get it out of my way and let my drivers work like they always used to; I want to disable the new thing."

From there it then also makes perfect self-consistent sense that "device_trust=full" would be the opposite end of the scale as in "Apply the explicit trust model 'fully', to all devices regardless of their bus-level default."

Now I guess you could name it something like "device.default_trust_level" to attempt to make it clearer that it represents a value the value represents the "trust" *of* each individual device itself, but TBH it's still liable to look backwards to many users - similarly it's hard to guess that FOO_POLICY_DISABLED might mean something entirely different from "don't enforce the policy for FOO.", thus rather than try to contort the naming of everything, I'd still be much more inclined to just flip the meanings to be more intuitive from the outsider perspective.

Say you upgrade your kernel and find your boot is broken because it now flibbles, but at least there is a "flibble" parameter whose values are "disabled", "wibble" and "frobble". Are you really going to spend time straight off reading up on what flibbling is to understand what wibbles and frobbles mean? Or is your instinct just going to be to try the seemingly obvious way of disabling flibbling, and then getting positively annoyed if that only makes it flibble harder?

Robin.