This code in agp_amd64_init() can optionally bind to devices generically, and it does this without going through the device core:
/* Look for any AGP bridge */
dev = NULL;
err = -ENODEV;
for_each_pci_dev(dev) {
if (!pci_find_capability(dev, PCI_CAP_ID_AGP))
continue;
/* Only one bridge supported right now */
if (agp_amd64_probe(dev, NULL) == 0) {
err = 0;
break;
}
}
This is causing crashes on some systems where sis-agp is already bound
to the device. It could even race with binding of another driver when
they are both built-in. I think it should be using something like the
following instead of directly calling agp_amd64_probe():
down(&dev->dev.sem);
if (!dev->dev.driver &&
driver_probe_device(&drv->driver, &dev->dev) == 0)
err = 0;
up(&dev->dev.sem);
if (err == 0)
break;
Ben.
--
Ben Hutchings
Power corrupts. Absolute power is kind of neat.
- John Lehman, Secretary of the US Navy 1981-1987
signature.asc
Description: This is a digitally signed message part
