On Sun, Jan 3, 2016 at 5:43 PM, Raymond Jennings shentino@gmail.com wrote:
At least the boot process gets all the information it needs (including the physical addresses associated with untyped memory frames) and also gets the master I/O capability.
On Sun, Jan 3, 2016 at 3:49 PM, Gerwin Klein Gerwin.Klein@nicta.com.au wrote:
On 26 Dec 2015, at 21:24, Corey Richardson corey@octayn.net wrote:
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.
The important part is that the only code that can manage and alter the IOMMU is trusted code.
https://en.wikipedia.org/wiki/Input%E2%80%93output_memory_management_unit
To say "you could put an arbitrary address" demands a clear understanding of who "you" is. User space device drivers are only possible if the user is trustworthy and only one instance of that user is allowed to touch the hardware registers or control block. IOMMU drivers are most common in video device drivers.
IOMMU registers are limited in many ways. PAT is a cleaner model for many things. Mixing PAT and IOMMU management is complex. It is hard to imagine a modern device and system without DMA today. The most interesting impact is processor IO is often limited to slow PIO transactions where DMA is an order of magnitude quicker or better on many systems. It is the rare IO card with enough memory to do interesting work without DMA.
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.
seL4 does not enforce absence of DMA. In general it’s not possible to do that, that’s why it’s an assumption to the proof that you need to validate for your system.
Without an IOMMU, you will need to trust the drivers and the hardware of DMA-capable devices to either not use DMA or to use it safely only.
You can lock down systems to not provide any access at all to DMA-devices, which sometimes is enough for simple separation-style systems. There are more trade-off points in the design space, but the trust story does become massively simpler and better when you have IOMMU support.
With an IOMMU, seL4 is in control and you’re fine.
Cheers, Gerwin
The information in this e-mail may be confidential and subject to legal professional privilege and/or copyright. National ICT Australia Limited accepts no liability for any damage caused by this email or its attachments.
Devel mailing list Devel@sel4.systems https://sel4.systems/lists/listinfo/devel