On Fri, Dec 25, 2015 at 11:19:55PM -0800, Raymond Jennings wrote:
> Speaking of which, how exactly does seL4 enforce no dma on systems without
> an IOMMU?
>
I've been unable to determine this myself. Consider the example of an ATA
controller: it seems you could put any arbitrary address in the PRDT and have
it spray disk contents into physical memory. In this case, all one needs is an
IOPort for that device's range on the IO bus and potentially for its place in
PCI configuration space (to enable bus mastering).
The manual seems to be self-contradictory here. In the section about the
BootInfo it indicates that the physical addresses are given to initiate DMA
when no IOMMU is present, but the IOSpace section states that to use DMA an
IOMMU must be used.
I can't really find any way that this is enforced or could be enforced by the
kernel. In userland it can be done quite easily by just not giving out
IOPorts.
> Also, what if the DMA controllers had its own device driver and the seL4
> kernel could be instructed to do the DMA buffer management itself and
> simply not allow any "userland" code to touch the DMA controller? Maybe if
> you have an untyped memory cap you can pass it to the kernel and ask it to
> be "mapped" to the dma controller.
>
Most DMA these days doesn't happen through any centralized "DMA controller",
but rather each device on the PCIe network can do DMA on its own if it is
enabled as a "bus master". This is why the IOMMU is critical - it provides a
single mechanism to enforce memory access policy when all sorts of devices
want to access memory, possibly without any other interaction with the host
CPU.