Skip disabled devices when enumerating ICDs via HKR

[manocharahul]

Problem

On Windows, ProbeDevice() in the HKR enumeration path only screens for reboot-pending states. A device that has been disabled in Device Manager is still reported as Valid.

Disabling a device does not remove its registry values, so the loader goes on to read OpenCLDriverName from the disabled adapter's HKR software key and registers an ICD for hardware that can never supply an OpenCL device.

The result is a platform with zero devices. That is legal, and clinfo handles it, but it is not useful — and applications that treat CL_DEVICE_NOT_FOUND as fatal fail to start even when another GPU in the system is present and working.

Why the HKR path specifically

Enumeration uses CM_GETIDLIST_FILTER_CLASS | CM_GETIDLIST_FILTER_PRESENT. In Windows terms present means physically installed — a disabled device is still present — so it is returned by CM_Get_Device_ID_ListW() and ProbeDevice() is the only gate after that.

The DXGK path is unaffected: it uses D3DKMTEnumAdapters2, which returns started adapters only.

Fix

Reject devices reporting CM_PROB_DISABLED so their ICD is never loaded. Both existing call sites already use if (ProbeDevice(...) != Valid) continue;, so no caller changes are needed, and CM_PROB_DISABLED comes from cfgmgr32.h, which is already included.

The check is deliberately narrow. It screens only the unambiguous, persistent case — a device a user explicitly disabled. CM_PROB_FAILED_START and a broader !(ulStatus & DN_STARTED) test were both considered and left out: the former can be transient (mid driver install, resource conflict), and ProbeDevice() is also called on software-component devnodes, where DN_STARTED semantics are less certain and a false rejection would hide a working ICD.

Verification

Reproduced on a two-GPU Windows system with one GPU disabled in Device Manager:

• Before: two platforms, one reporting 0 devices. Loader trace shows the disabled device accepted and its ICD path read.
• After: one platform with the working GPU. Trace shows WARNING: device is disabled (0x1802400), skipping..., and the disabled device's OpenCLDriverName is never read or its ICD loaded.

clinfo exits 0 in both cases; the difference is that the useless empty platform is no longer advertised.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[jenatali]

Seems reasonable to me.

[manocharahul]

I have signed CLA but its still not updating. Can someone please check.

[bashbaug]

I have signed CLA but its still not updating. Can someone please check.

You need to address this:

Rahul Manocha seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.

[manocharahul]

I added my email to my profile and commit shows verified user now. Signed the CLA again, but doesn't seem to get updated

[manocharahul]

CLA is signed now. Can we get workflow approval?

[manocharahul]

@bashbaug How do i get the workflow approved?

[bashbaug]

Merging as discussed in the September 22nd teleconference.