TrustZone: transfering information cross-world using interrupts

Martin Stein martin.stein at ...1...
Mon Nov 9 14:31:48 CET 2015

Hi David,

Am 06.11.2015 um 15:39 schrieb David Goltzsche:
> One possible approach would be to write something to a certain location
> in the non-secure memory before triggering the interrupt and reading
> that location from Linux. Of course, a cross-world synchronization
> mechanism is required here to maintain the consistency of the reserved
> memory range.

I would recommend this approach. An example how to implement this, is
the paravirtualized block-driver for the Linux-VM on my working branches
1497_usb_armory_demo [1] and [2]. The secure back-end of the driver is
implemented in [3] whereas the unsecure front-end can be found in [4].
Initially, genode_blk_init (Linux) installs a piece of unsecured RAM as
communication buffer by requesting Vmm::Block::_buffer (Genode). The
latter checks that the buffer is completely in the VMs RAM and
determines its local base. Now, as soon as the device signals the
completion of a block request to the VMM, Callback::entry (Genode)
injects an IRQ into the VM. Then, event_interrupt (Linux) requests
Vmm::Block::_collect_reply (Genode) which writes the result of the block
request into the communictaion buffer. For proper communication, it is
important that the buffer is mapped uncached on both secure and unsecure
side. On the unsecure side, this is achieved via dma_alloc_coherent wich
unfortunately requires a device object as input. As we have no real
device, we create a dummy device object for that purpose at the
beginning of genode_blk_init. On the secure side, the whole VM RAM is
mapped as "IO-MEM", which is uncached by default (see _ram and
_ram_iomem in [5]).

> Another approach could use Linux's shared interrupts: Additional
> information can be transferred via the dev_id pointer passed to an
> interrupt handler.

Admittedly, I have don't know whether this approach is feasible and, in
case it is, how much effort it would be to implement it. But I assume
that communicating via shared RAM means less effort and higher flexibility.


[3] <GENODE>/repos/os/src/server/tz_vmm/
[4] <LINUX>/drivers/block/genode.c
[5] <GENODE>/repos/os/src/server/tz_vmm/include/vm_base.h

More information about the users mailing list