Re: [RFC PATCH 1/4] bus: fsl-mc: add custom .dma_configure implementation

From: Robin Murphy
Date: Wed Apr 15 2020 - 13:32:08 EST

On 2020-04-15 4:44 pm, Laurentiu Tudor wrote:


On 4/14/2020 5:32 PM, Lorenzo Pieralisi wrote:
On Wed, Mar 25, 2020 at 06:48:55PM +0200, Laurentiu Tudor wrote:
Hi Lorenzo,

On 3/25/2020 2:51 PM, Lorenzo Pieralisi wrote:
On Thu, Feb 27, 2020 at 12:05:39PM +0200, laurentiu.tudor@xxxxxxx wrote:
From: Laurentiu Tudor <laurentiu.tudor@xxxxxxx>

The devices on this bus are not discovered by way of device tree
but by queries to the firmware. It makes little sense to trick the
generic of layer into thinking that these devices are of related so
that we can get our dma configuration. Instead of doing that, add
our custom dma configuration implementation.

Signed-off-by: Laurentiu Tudor <laurentiu.tudor@xxxxxxx>
---
drivers/bus/fsl-mc/fsl-mc-bus.c | 31 ++++++++++++++++++++++++++++++-
1 file changed, 30 insertions(+), 1 deletion(-)

diff --git a/drivers/bus/fsl-mc/fsl-mc-bus.c b/drivers/bus/fsl-mc/fsl-mc-bus.c
index 36eb25f82c8e..eafaa0e0b906 100644
--- a/drivers/bus/fsl-mc/fsl-mc-bus.c
+++ b/drivers/bus/fsl-mc/fsl-mc-bus.c
@@ -132,11 +132,40 @@ static int fsl_mc_bus_uevent(struct device *dev, struct kobj_uevent_env *env)
static int fsl_mc_dma_configure(struct device *dev)
{
struct device *dma_dev = dev;
+ struct iommu_fwspec *fwspec;
+ const struct iommu_ops *iommu_ops;
+ struct fsl_mc_device *mc_dev = to_fsl_mc_device(dev);
+ int ret;
+ u32 icid;
while (dev_is_fsl_mc(dma_dev))
dma_dev = dma_dev->parent;
- return of_dma_configure(dev, dma_dev->of_node, 0);
+ fwspec = dev_iommu_fwspec_get(dma_dev);
+ if (!fwspec)
+ return -ENODEV;
+ iommu_ops = iommu_ops_from_fwnode(fwspec->iommu_fwnode);
+ if (!iommu_ops)
+ return -ENODEV;
+
+ ret = iommu_fwspec_init(dev, fwspec->iommu_fwnode, iommu_ops);
+ if (ret)
+ return ret;
+
+ icid = mc_dev->icid;
+ ret = iommu_fwspec_add_ids(dev, &icid, 1);

I see. So with this patch we would use the MC named component only to
retrieve the iommu_ops

Right. I'd also add that the implementation tries to follow the existing
standard .dma_configure implementations, e.g. of_dma_configure +
of_iommu_configure. I'd also note that similarly to the ACPI case, this
MC FW device is probed as a platform device in the DT scenario, binding
here [1].
A similar approach is used for the retrieval of the msi irq domain, see
following patch.

- the streamid are injected directly here bypassing OF/IORT bindings translations altogether.

Actually I've submitted a v2 [2] that calls into .of_xlate() to allow
the smmu driver to do some processing on the raw streamid coming from
the firmware. I have not yet tested this with ACPI but expect it to
work, however, it's debatable how valid is this approach in the context
of ACPI.

Actually, what I think you need is of_map_rid() (and an IORT
equivalent, that I am going to write - generalizing iort_msi_map_rid()).

Would that be enough to enable IORT "normal" mappings in the MC bus
named components ?


At a first glance, looks like this could very well fix the ACPI
scenario, but I have some unclarities on the approach:
* are we going to rely in DT and ACPI generic layers even if these
devices are not published / enumerated through DT or ACPI tables?

Assuming you mean the DPRC devices rather than the MC itself, then yes; in that sense it's exactly the same as how we treat dynamically-discovered PCI devices.

* the firmware manages and provides discrete streamids for the devices
it exposes so there's no translation involved. There's no
requestor_id / input_id involved but it seems that we would still do
some kind of translation relying for this on the DT/ACPI functions.

Wrong - last time I looked, what that firmware actually manages are *ICIDs* for the devices, not SMMU Stream IDs or GIC Device IDs; what DT/ACPI specifies is a translation from ICID to Stream ID/Device ID. The ICID is very much the requester/input ID for that translation. Yes, in practice the "translation" is effectively always a trivial identity mapping, but conceptually it most definitely exists. Yes, the subtlety is incredibly easy to overlook because it's basically drawing a distinction between one end of some wires vs. the other end, but it matters.

(and of course "trivial 1:1 translation" isn't even true in the case of SMMU Stream ID values, since IIRC they are really composed of 5 different inputs, only one of which is (part of) the incoming ICID)

* MC firmware has its own stream_id (e.g. on some chips 0x4000, others
0xf00, so outside the range of stream_ids used for the mc devices)
while for the devices on this bus, MC allocates stream_ids from a
range (e.g. 0x17 - 0x3f). Is it possible to describe this in the IORT table?

If it represents a unique ICID allocated to the MC itself, then sure, it simply goes through the mapping like anything else. Just like a PCI host bridge owns requester ID 0:0.0 and thus whatever Stream ID/Device ID that might map to.

If (for the sake of argument, because AIUI everything is an ICID in this particular case) it's some hard-wired thing that exists in Stream ID/Device ID space only, then it's a little trickier, but still in scope. In DT we have a lovely distinction between between "originating from the node" and "translated through the node", e.g. "msi-parent" vs. "msi-map"; IORT is not quite as clear-cut, but there are at least a few options. If the valid input ID space is smaller than 32 bits, then the "Named Component as bridge" binding could simply define special out-of-range values to represent IDs originating from the bridge itself, such that the NC driver knows what to do and from IORT's point of view everything is just a normal mapping. Alternatively there's already the example of SMMUv3 where we can have a mix of the normal mappings from Stream ID to Device ID for the upstream masters plus a single mapping for the SMMU's own Device ID - admittedly that depends on the additional SMMUv3-specific Device ID Mapping Index property, but if necessary it might be workable to have a de-facto interface for NCs that only considers single mappings when configuring the NC itself, and only considers normal mappings when configuring its children. Or maybe define a new mapping flag or NC property if there's a real need to specify such a situation unambiguously at the IORT level.

* Regarding the of_map_rid() use you mentioned, I was planning to
decouple the mc bus from the DT layer by dropping the use of
of_map_rid(), see patch 4.
I briefly glanced over the iort code and spotted this static function:
iort_iommu_xlate(). Wouldn't it also help, of course after making it public?

I won't speak for Lorenzo or claim we've agreed on an approach at all (not least because in all honesty we haven't really discussed it beyond these various email threads), but FWIW my vision is that ultimately the DT/ACPI code would expose a *_dma_configure() interface that takes an optional input ID, or (perhaps more likely) an explicit pair of interfaces for "configure this regular device" and "configure this device based on this 'host' device and this ID", and it becomes the bus code's responsibility to pass the right device(s) and deal with multiple IDs (i.e. for starters all the PCI alias stuff goes back to the PCI code where it really should be, rather than having multiple copies of magic PCI awareness deep down in DT/ACPI code).

Robin.