I tracked this problem down to a recent optimization made in the Linux implementation of os_pci_init_handle(). The attached patch fixes this problem. You can apply it using:
# sh /path/to/NVIDIA-Linux-x86-177.68-pkg0.run \
--apply-patch /path/to/NVIDIA_kernel-177.68-2714327.diff.txt
# sh NVIDIA-Linux-x86-177.68-pkg0-custom.run