[Syllo]

Summary

Intel GPUs are discarded during enumeration when /sys/bus/pci/devices/<pci>/enable is not exactly "1":

if (nvtop_device_get_sysattr_value(parent, "enable", &enabled) < 0 || strcmp(enabled, "1"))
  return;

That attribute is not a boolean. It exposes pci_dev.enable_cnt, the number of times pci_enable_device() has been called:

• drivers/pci/pci-sysfs.c: enable_show() emits atomic_read(&pdev->enable_cnt).
• drivers/pci/pci.c: pci_enable_device_flags() does atomic_inc_return(&dev->enable_cnt).
• include/linux/pci.h: pci_is_enabled() is enable_cnt > 0.

Integrated GPUs commonly have a count greater than one. On the affected Meteor Lake-P machine it is 3, so nvtop never registers the GPU and prints No GPU to monitor.

The value is also not a stable indicator of usability: xe_pci_runtime_suspend() and xe_pci_suspend() call pci_disable_device() for D3cold while the driver remains bound, so the count can be zero for a perfectly functional, idle GPU.

After the existing Intel vendor and i915/xe driver checks, the bound driver is the reliable signal: both drivers enable the device during probe (i915_driver_probe() calls pci_enable_device(), xe_pci_probe() calls pcim_enable_device()), and probe failure prevents binding. This removes the redundant and fragile enable check.

Testing

• Built with -DINTEL_SUPPORT=ON.
• All 27 existing tests pass.
• nvtop -s still works normally on the non-Intel GPU of the development machine.
• Cannot test on affected Intel hardware directly; the reporter confirmed the one-line variant fixes detection on their Meteor Lake-P system.

Fixes #553
See also #465, #498 (not confirmed to share this root cause).