Hi Stephen,
On 17/09/18 22:36, Stephen Warren wrote:
Joerg, Christoph, Marek, Robin,
I believe that the driver for our PCIe endpoint controller hardware will
need to explicitly manage its IOVA space more than current APIs allow.
I'd like to discuss how to make that possible.
First some background on our hardware:
NVIDIA's Xavier SoC contains a Synopsis Designware PCIe controller. This
can operate in either root port or endpoint mode. I'm particularly
interested in endpoint mode.
Our particular instantiation of this controller exposes a single
function with a single software-controlled PCIe BAR to the PCIe bus
(there are also BARs for access to DMA controller registers and outbound
MSI configuration, which can both be enabled/disabled but not used for
any other purpose). When a transaction is received from the PCIe bus,
the following happens:
1) Transaction is matched against the BAR base/size (in PCIe address
space) to determine whether it "hits" this BAR or not.
2) The transaction's address is processed by the PCIe controller's ATU
(Address Translation Unit), which can re-write the address that the
transaction accesses.
Our particular instantiation of the hardware only has 2 entries in the
ATU mapping table, which gives very little flexibility in setting up a
mapping.
As an FYI, ATU entries can match PCIe transactions either:
a) Any transaction received on a particular BAR.
b) Any transaction received within a single contiguous window of PCIe
address space. This kind of mapping entry obviously has to be set up
after device enumeration is complete so that it can match the correct
PCIe address.
Each ATU entry maps a single contiguous set of PCIe addresses to a
single contiguous set of IOVAs which are passed to the IOMMU.
Transactions can pass through the ATU without being translated if desired.
3) The transaction is passed to the IOMMU, which can again re-write the
address that the transaction accesses.
4) The transaction is passed to the memory controller and reads/writes
DRAM.
In general, we want to be able to expose a large and dynamic set of data
buffers to the PCIe bus; certainly /far/ more than two separate buffers
(the number of ATU table entries). With current Linux APIs, these
buffers will not be located in contiguous or adjacent physical (DRAM) or
virtual (IOVA) addresses, nor in any particular window of physical or
IOVA addresses. However, the ATU's mapping from PCIe to IOVA can only
expose one or two contiguous ranges of IOVA space. These two sets of
requirements are at odds!
So, I'd like to propose some new APIs that the PCIe endpoint driver can
use:
1) Allocate/reserve an IOVA range of specified size, but don't map
anything into the IOVA range.
2) De-allocate the IOVA range allocated in (1).
3) Map a specific set (scatter-gather list I suppose) of
already-allocated/extant physical addresses into part of an IOVA range
allocated in (1).
4) Unmap a portion of an IOVA range that was mapped by (3).
That all sounds perfectly reasonable - basically it sounds like the
endpoint framework wants the option to do the same as VFIO or many DRM
drivers, i.e. set up its own IOMMU domain, attach the endpoint's group,
and explicitly manage its mappings via IOMMU API calls. Provided you can
assume cache-coherent PCI, that should be enough to get things going -
supporting non-coherent endpoints is a little trickier in terms of
making sure the endpoint controller and/or device gets the right DMA ops
to only ever perform cache maintenance once you add streaming DMA
mappings into the mix, but that's not insurmountable (and I think it's
something we still need to address for DRM anyway, at least on arm64)
One final note:
The memory controller can translate accesses to a small region of DRAM
address space into accesses to an interrupt generation module. This
allows devices attached to the PCIe bus to generate interrupts to
software running on the system with the PCIe endpoint controller. Thus I
deliberately described API 3 above as mapping a specific physical
address into IOVA space, as opposed to mapping an existing DRAM
allocation into IOVA space, in order to allow mapping this interrupt
generation address space into IOVA space. If we needed separate APIs to
map physical addresses vs. DRAM allocations into IOVA space, that would
likely be fine too.
If that's the standard DesignWare MSI dingaling, then all you should
need to do is ensure you IOVA is reserved in your allocator (if it can
be entirely outside the EP BAR, even better) - AFAIK the writes get
completely intercepted such that they never go out to the SMMU side at
all, and thus no actual mapping is even needed.
Does this API proposal sound reasonable?
Indeed, as I say apart from using streaming DMA for coherency management
(which I think could be added in pretty much orthogonally later), this
sounds like something you could plumb into the endpoint framework right
now with no dependent changes elsewhere.
I have heard from some NVIDIA developers that the above APIs rather go
against the principle that individual drivers should not be aware of the
presence/absence of an IOMMU, and hence direct management of IOVA
allocation/layout is deliberately avoided, and hence there hasn't been a
need/desire for this kind of API in the past. However, I think our
current hardware design and use-case rather requires it. Do you agree?
If there is a principle, it's more the inverse - the point of things
like SWIOTLB and iommu-dma is that we don't want to *have* to add
IOMMU-awareness or explicit bounce-buffering to every driver or
subsystem which might ever find itself on a machine with more memory
than its device can address natively. Thus drivers which only need to
use the DMA API can continue to do so and the arch code hooks up this
stuff automatically to make sure that just works. However, drivers which
*do* expect their device to have an IOMMU, and have good cause to manage
it themselves to do things that simple DMA API calls can't, should of
course be welcome to implement that extra code and depend on IOMMU_API
if they so wish. Again, DRM drivers are the prime example (er, no pun
intended) - simple ones let drm_gem_cma_helper et al do all the heavy
lifting for them, more complex ones get their hands dirty.
Robin.
