Freeze during "initialising devices" in Virtualbox when chainloading

[mcb30]

Same problem with Virtualbox 7.0.8 and all Intel 1000 cards in EFI mode

Originally posted by @Reonaydo in #624 (comment)

Reproduction instructions: create new VirtualBox VM with EFI mode selected, attempt chainloading ipxe.efi.

[mcb30]

Testing with some extra debug shows that the freeze occurs when we attempt to open the PCI I/O protocol with BY_DRIVER attributes:

	/* Open PCI device */
	if ( ( rc = efipci_open ( device, ( EFI_OPEN_PROTOCOL_BY_DRIVER |
					    EFI_OPEN_PROTOCOL_EXCLUSIVE ),
				  efipci ) ) != 0 ) {
		DBGC ( device, "EFIPCI %s could not open PCI device: %s\n",
		       efi_handle_name ( device ), strerror ( rc ) );
		DBGC_EFI_OPENERS ( device, device, &efi_pci_io_protocol_guid );
		goto err_open;
	}

The debug messages show that even after disconnection, the VirtualBox "E1000 network interface card Driver" still has the handle open with BY_DRIVER attributes. This suggests that the freeze is caused by the platform firmware being stuck in a loop attempting to repeatedly disconnect the driver, but the driver is failing each time to close its BY_DRIVER handle.

[mcb30]

This failure of the VirtualBox "E1000 network interface card Driver" to disconnect properly can be reproduced without any iPXE code using the UEFI shell:

[mcb30]

I have found the root cause bug within the VirtualBox VBoxPkg E1kNetDxe driver.

The VirtualBox driver's Stop() method will do absolutely nothing unless called with a non-zero number of children:

STATIC
EFI_STATUS
EFIAPI
E1kNetDriverBindingStop (
  IN EFI_DRIVER_BINDING_PROTOCOL *This,
  IN EFI_HANDLE                  ControllerHandle,
  IN UINTN                       NumberOfChildren,
  IN EFI_HANDLE                  *ChildHandleBuffer
  )
{
  if (NumberOfChildren > 0) {
    //
    // free all resources for whose access we need the child handle, because
    // the child handle is going away
    //
    EFI_STATUS                  Status;
    EFI_SIMPLE_NETWORK_PROTOCOL *Snp;
    E1K_NET_DEV                 *Dev;
    EFI_TPL                     OldTpl;

    ASSERT (NumberOfChildren == 1);

    ......

      gBS->CloseProtocol (
             ControllerHandle,
             &gEfiPciIoProtocolGuid,
             This->DriverBindingHandle,
             ControllerHandle
             );

    ......
  }

  return EFI_SUCCESS;
}

This relies on the driver's Start() method having opened the underlying PCI I/O protocol handle with BY_CHILD_CONTROLLER attributes on behalf of its newly created SNP protocol handle, so that the UEFI firmware can be aware of the parent-child relationship and therefore pass in a list of child handles to the Stop() method. However, the VirtualBox driver's Start() method never performs this BY_CHILD_CONTROLLER open.

This bug can be observed directly in the UEFI shell's list of devices:

Note that device AB (the SNP protocol handle) is not shown as a child of device 9E (the PCI I/O protocol handle).

The net effect is that the VirtualBox driver's Stop() method is a pure no-op: it will never close the PCI I/O protocol handle as required.

[veeyoung]

solved it by chaging network adapter type from intel pro 1000 to pcnet fast