#!/bin/python3
#
# vGPU unlock script for consumer GPUs.
#
# Copyright 2021 Jonathan Johansson
# This file is part of the "vgpu_unlock" project, and is distributed under the MIT License.
# See the LICENSE file for more details.
#


import errno
import frida
import os
import queue
import subprocess
import sys
import time

script_source = r"""
    // Value of the "request" argument used by nvidia-vgpud and nvidia-vgpu-mgr
    // when calling ioctl to read the PCI device ID and type (and possibly
    // other things) from the GPU.
    var REQ_QUERY_GPU = ptr("0xC020462A");

    // When issuing ioctl with REQ_QUERY_GPU then the "argp" argument is a
    // pointer to a structure something like this:
    //
    // struct arg {
    //    uint32_t unknown_1; // Initialized prior to call.
    //    uint32_t unknown_2; // Initialized prior to call.
    //    uint32_t op_type;   // Operation type, see comment below.
    //    uint32_t padding_1; // Always set to 0 prior to call.
    //    void*    result;    // Pointer initialized prior to call.
    //                        // Pointee initialized to 0 prior to call.
    //                        // Pointee is written by ioctl call.
    //    uint32_t unknown_4; // Set to 0x10 for READ_PCI_ID and set to 4 for
                              // READ_DEV_TYPE prior to call.
    //    uint32_t status;    // Written by ioctl call. See comment below.
    // }

    // These are the observed values for the op_type member.
    var OP_READ_DEV_TYPE = 0x800289; // *result type is uint64_t.
    var OP_READ_PCI_ID = 0x20801801; // *result type is uint16_t[4], the second
                                     // element (index 1) is the device ID, the
                                     // forth element (index 3) is the subsystem
                                     // ID.

    // nvidia-vgpu-mgr expects this value for a vGPU capable GPU.
    var DEV_TYPE_VGPU_CAPABLE = uint64(3);

    // When ioctl returns success (retval >= 0) but sets the status value of
    // the arg structure to 3 then nvidia-vgpud will sleep for a bit (first
    // 0.1s then 1s then 10s) then issue the same ioctl call again until the
    // status differs from 3. It will attempt this for up to 24h before giving
    // up.
    var STATUS_TRY_AGAIN = 3;

    Interceptor.attach(Module.getExportByName(null, "ioctl"), {
        onEnter(args) {
            this.request = args[1];
            this.argp = args[2];
        },
        onLeave(retVal) {
            if(!this.request.equals(REQ_QUERY_GPU)) {
                // Not a call we care about.
                return;
            }

            if(retVal.toInt32() < 0) {
                // Call failed.
                return;
            }

            // Lookup status value according to struct above.
            var status = this.argp.add(0x1C).readU32();

            if(status == STATUS_TRY_AGAIN) {
                // Driver will try again.
                return;
            }

            var op_type = this.argp.add(8).readU32();

            if(op_type == OP_READ_PCI_ID) {
                // Lookup address of the device and subsystem IDs.
                var devid_ptr = this.argp.add(0x10).readPointer().add(2);
                var subsysid_ptr = this.argp.add(0x10).readPointer().add(6);

                // Now we replace the device ID with a spoofed value that needs to
                // be determined such that the spoofed value represents a GPU with
                // vGPU support that uses the same GPU chip as our actual GPU.
                var actual_devid = devid_ptr.readU16();
                var spoofed_devid = actual_devid;
                var actual_subsysid = subsysid_ptr.readU16();
                var spoofed_subsysid = actual_subsysid;

                // GP102
                if(actual_devid == 0x1b00 || // TITAN X (Pascal)
                   actual_devid == 0x1b02 || // TITAN Xp
                   actual_devid == 0x1b06 || // GTX 1080 Ti
                   actual_devid == 0x1b30) { // Quadro P6000
                    spoofed_devid = 0x1b38; // Tesla P40
                }

                // GP104
                if(actual_devid == 0x1b80 || // GTX 1080
                   actual_devid == 0x1b81 || // GTX 1070
                   actual_devid == 0x1b82 || // GTX 1070 Ti
                   actual_devid == 0x1b83 || // GTX 1060 6GB
                   actual_devid == 0x1b84 || // GTX 1060 3GB
                   actual_devid == 0x1bb0) { // Quadro P5000
                    spoofed_devid = 0x1bb3; // Tesla P4
                }

                // TU102
                if(actual_devid == 0x1e02 || // TITAN RTX
                   actual_devid == 0x1e04 || // RTX 2080 Ti
                   actual_devid == 0x1e07) { // RTX 2080 Ti Rev. A
                    spoofed_devid = 0x1e30; // Quadro RTX 6000
                    spoofed_subsysid = 0x12ba;
                }

                // TU104
                if(actual_devid == 0x1e81 || // RTX 2080 Super
                   actual_devid == 0x1e82 || // RTX 2080
                   actual_devid == 0x1e84 || // RTX 2070 Super
                   actual_devid == 0x1e87 || // RTX 2080 Rev. A
                   actual_devid == 0x1e89 || // RTX 2060
                   actual_devid == 0x1eb0 || // Quadro RTX 5000
                   actual_devid == 0x1eb1) { // Quadro RTX 4000
                    spoofed_devid = 0x1eb8; // Tesla T4
                }

                // GA102
                if(actual_devid == 0x2204 || // RTX 3090
                   actual_devid == 0x2205 || // RTX 3080 Ti
                   actual_devid == 0x2206) { // RTX 3080
                    spoofed_devid = 0x2235; // RTX A40
                }

                devid_ptr.writeU16(spoofed_devid);
                subsysid_ptr.writeU16(spoofed_subsysid);
            }
            
            if(op_type == OP_READ_DEV_TYPE) {
                // Set device type to vGPU capable.
                var dev_type_ptr = this.argp.add(0x10).readPointer();
                dev_type_ptr.writeU64(DEV_TYPE_VGPU_CAPABLE);
            }
        }
    });
"""

device = frida.get_local_device()
child_processes = queue.Queue()

def instrument(pid):
    """Instrument and resume process.

    :param pid: Process identifier
    """

    session = device.attach(pid)
    # We need to also instrument the children since nvidia-vgpud forks itself
    # when initially launched.
    session.enable_child_gating()
    script = session.create_script(script_source)
    script.load()
    device.resume(pid)


def on_child_added(child):
    """Callback for when a new child process has been created.

    :param child: The newly created child process.
    """

    child_processes.put(child.pid)
    instrument(child.pid)


def wait_exit(pid):
    """Wait for a process to terminate.

    :param pid: Process ID of the target process.
    """

    while 1:
        time.sleep(.1)

        try:
            os.kill(pid, 0)

        except OSError as e:
            if e.errno == errno.ESRCH:
                break


def main():
    """Entrypoint."""

    # Behave at least a little bit like a forking service.
    if sys.argv[1] != "-f":
        subprocess.Popen([sys.argv[0], "-f"] + sys.argv[1:])
        exit()

    device.on("child-added", on_child_added)
    pid = device.spawn(sys.argv[2])
    instrument(pid)

    # Wait for everything to terminate before exiting.
    wait_exit(pid)

    while not child_processes.empty():
        wait_exit(child_processes.get_nowait())


if __name__ == "__main__":
    main()