I'd prefer to pass the descriptor as a parameter rather than introducing a global variable.

Could maybe s/def/default/. :)

Should we also check the device class and subclass?

or "prove"

Please limit capabilities on the descriptor, see the use of caph_rights_limit() and caph_ioctls_limit() in passthru_init(). Actually I do not see why the descriptor needs to be kept open indefinitely.

Please add a comment explaining why we look at the host device.

The way this is handled in pci_hostbridge.c might be a bit simpler than needing set_config_value_if_unset, etc. It does something like this:

u_int vendor, device, revid, subvendor, subdevice;
const char *value;

vendor = LPC_DEV;
device = 0;
...

/* On Intel systems... */
if (....) {
    vendor = read_config(&sel, PCIR_VENDOR, 2);
    ...
}

/* Check for config variable overrides. */
value = get_config_value_node(nvl, "vendor");
if (value != NULL)
   vendor = strtol(value, NULL, 0);
...

It might also be nice to use a single PCIOCGETCONF request to fetch the 'conf' structure for the LPC rather than a bunch of read_config ioctls. I'm also still a bit torn on if we need a separate 'lpc.use_host_ids' or the like that gated the if clause to override the values with host values.

Sorry, I was not clear. It's not really worthwhile to limit rights if the descriptor is not kept open. Since it's closed right after this, the caph_* calls can be dropped IMO.

In general I'd like some way to not force this on always perhaps? Or at least provide a configuration knob to allow it to be turned off in case it breaks other systems that don't need GVT-d.

Ideally I think what I would like is for the lpc device to honor config values to set these registers and have a hook here that sets the config variables if they aren't always set (so the user can always override them if needed).