21.01.2020 23:12, Dmitry Osipenko пишет:
> 21.01.2020 18:54, Thierry Reding пишет:
>> On Tue, Jan 21, 2020 at 05:18:43PM +0200, Georgi Djakov wrote:
>>> On 1/21/20 16:10, Thierry Reding wrote:
>>>> On Tue, Jan 21, 2020 at 09:53:48AM +0300, Dmitry Osipenko wrote:
>>>>> 20.01.2020 18:06, Thierry Reding пишет:
>>>>>> On Fri, Jan 17, 2020 at 05:23:43PM +0200, Georgi Djakov wrote:
>>>>>>> Hi Thierry,
>>>>>>>
>>>>>>> Thanks for the patch!
>>>>>>>
>>>>>>> On 1/14/20 20:15, Thierry Reding wrote:
>>>>>>>> From: Thierry Reding <treding@xxxxxxxxxx>
>>>>>>>>
>>>>>>>> Document the interconnects property that is used to describe the paths
>>>>>>>> from and to system memory from and to the BPMP.
>>>>>>>>
>>>>>>>> Signed-off-by: Thierry Reding <treding@xxxxxxxxxx>
>>>>>>>> ---
>>>>>>>> Rob, Georgi,
>>>>>>>>
>>>>>>>> after the initial RFC that I did for adding interconnect properties on
>>>>>>>> Tegra, I realized that the description wasn't complete. This is an
>>>>>>>> attempt at a more accurate description, but unfortunately I'm not sure
>>>>>>>> if it's even correct in terms of the interconnect bindings.
>>>>>>>>
>>>>>>>> The problem here is that on Tegra, each device has multiple paths to
>>>>>>>> system memory, and I have no good idea on what to pick as the default.
>>>>>>>> They are all basically the same path, but each provides extra controls
>>>>>>>> to configure the "interconnect".
>>>>>>>
>>>>>>> Are these multiple paths between a device and system memory used simultaneously
>>>>>>> for load-balancing, or who makes the decision about which path would be used?
>>>>>>
>>>>>> It varies. The vast majority of these paths are read/write pairs, which
>>>>>> can be configured separately. There are also cases where multiple paths
>>>>>> are used for load-balancing and I don't think there's any direct
>>>>>> software control over which path will be used.
>>>>>>
>>>>>> A third class is where you have one device, but two read/write pairs,
>>>>>> one which is tied to a microcontroller that's part of the device, and
>>>>>> another read/write pair that is used for DMA to/from the device.
>>>>>>
>>>>>> Often in the latter case, the microcontroller memory client interfaces
>>>>>> will be used by the microcontroller to read firmware and once the micro-
>>>>>> controller has booted up, the DMA memory client interfaces will be used
>>>>>> to read/write system memory with bulk data (like frame buffers, etc.).
>>>>>>
>>>>>>> Is this based on the client/stream ID that you mentioned previously?
>>>>>>
>>>>>> These are now all what's call memory client IDs, which identify the
>>>>>> corresponding interface to the memory controller. Stream IDs are
>>>>>> slightly higher-level and typically identify the "module" that uses
>>>>>> the SMMU. Generally a stream ID is mapped to one or more memory client
>>>>>> IDs.
>>>>>>
>>>>>>> Looking at the the binding below, it seems to me like there are different
>>>>>>> master/slave pairs between MC and EMC and each link is used for
>>>>>>> unidirectional traffic only. In terms of the interconnect API, both read
>>>>>>> and write paths have the same direction.
>>>>>
>>>>> Yes, that definition should be incorrect.
>>>>>
>>>>>> I'm not sure I understand what you mean by this last sentence. Are you
>>>>>> saying that each path in terms of the interconnect API is a always a
>>>>>> bidirectional link?
>>>>>
>>>>> Please see more below.
>>>>>
>>>>>>> Is the EMC really an interconnect provider or is it just a slave port? Can
>>>>>>> we scale both EMC and MC independently?
>>>>>>
>>>>>> The EMC is the only one where we can scale the frequency, but the MC has
>>>>>> various knobs that can be used to fine-tune arbitration, set maximum
>>>>>> latency, etc.
>>>>>
>>>>> Yes..
>>>>>
>>>>>
>>>>> EMC controls the total amount of available memory bandwidth, things like
>>>>> DRAM timing and EMC-DRAM channel's performance. EMC is facing MC from
>>>>> one side and DRAM (EMEM) from the other.
>>>>>
>>>
>>> Right, so we can use the icc framework here to aggregate the requested bandwidth
>>> from all clients and scale the frequency/voltage of EMC.
>>
>> Yeah, that was the plan. Dmitry's patches implement most of that. Note
>> that we're working on this from two sides: on one hand we need to figure
>> out the bindings so that we can set up the interconnect paths, then the
>> memory controller and external memory controller drivers need to be made
>> interconnect providers so that we can actually implement the dynamic
>> scaling (and then finally all memory client drivers need to be updated
>> to actually use these ICC framework to request bandwidth).
>>
>> I'm prioritizing the first issue right now because that's currently a
>> blocker for enabling SMMU support on Tegra194 where we need to set the
>> DMA mask based on the "bus" (i.e. DMA parent, i.e. the MC) because there
>> are additional restrictions that don't exist on prior generations where
>> we can simply set the DMA mask to the device's "native" DMA mask.
>>
>> EMC frequency scaling is slightly lower on my priority list because in
>> most use-cases there is enough bandwidth at the default EMC frequency.
>>
>>>>> MC controls allocation of that total bandwidth between the memory
>>>>> clients. It has knobs to prioritize clients, the knobs are per
>>>>> read/write port. MC is facing memory clients from one side and EMC from
>>>>> the other.
>>>>>
>>>
>>> Thanks for clarifying! So are these QoS knobs (priority, latency etc.) tuned
>>> dynamically during runtime or is it more like a static configuration that is
>>> done for example just once during system boot?
>>
>> The hardware defaults are usually sufficient unless the system is under
>> very high memory pressure. I'm only aware of one case where we actually
>> need to override the hardware default on boot and it's exotic enough to
>> not be upstream yet.
>>
>> Ultimately we want to tune these at runtime, typically together with and
>> depending on the EMC frequency. Under memory pressure you can use the
>> "latency allowance" registers to prioritize memory requests from
>> isochronous clients (like display controllers) to ensure they don't
>> underrun.
>
> Perhaps USB could be one example of a memory client that may need ISO
> transfers for multimedia things (live audio/video), while ISO not needed
> for file transfers.
>
>> Given that we only have to tune these under somewhat extreme conditions,
>> I think these are lower priority from an implementation point of view
>> than the EMC frequency scaling, but the registers are based on the
>> memory client IDs, so I think it's convenient to have those be part of
>> the bindings in the first place.
>>
>>>>>> I vaguely recall Dmitry mentioning that the EMC in early generations of
>>>>>> Tegra used to have controls for individual memory clients, but I don't
>>>>>> see that in more recent generations.
>>>>>
>>>>> EMC doesn't have direct controls over memory clients on all Tegra SoCs,
>>>>> but it may have some extra knobs for the MC arbitration config.
>>>>>
>>>>> The MC bandwidth allocation logic and hardware programming interface
>>>>> differs among SoC generations, but the basic principle is the same.
>>>>>
>>>>>>>> Any ideas on how to resolve this? Let me know if the DT bindings and
>>>>>>>> example don't make things clear enough.
>>>>>
>>>>> I'm also interested in the answer to this question.
>>>>>
>>>>> A quick thought.. maybe it could be some new ICC DT property which tells
>>>>> that all paths are the "dma-mem":
>>>>>
>>>>> interconnects-all-dma-mem;
>>>>
>>>> There could easily be cases where multiple interconnects are to system
>>>> memory but there are additional ones which aren't, so the above wouldn't
>>>> be able to represent such cases.
>>>
>>> Yes, true.
>
> Sure, but then that property shouldn't be used.
>
> Anyways, I think Thierry's suggestion about the generic memory
> controller API sounds more attractive.
>
>>>>>>>> .../firmware/nvidia,tegra186-bpmp.yaml | 59 +++++++++++++++++++
>>>>>>>> 1 file changed, 59 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/Documentation/devicetree/bindings/firmware/nvidia,tegra186-bpmp.yaml b/Documentation/devicetree/bindings/firmware/nvidia,tegra186-bpmp.yaml
>>>>>>>> index dabf1c1aec2f..d40fcd836e90 100644
>>>>>>>> --- a/Documentation/devicetree/bindings/firmware/nvidia,tegra186-bpmp.yaml
>>>>>>>> +++ b/Documentation/devicetree/bindings/firmware/nvidia,tegra186-bpmp.yaml
>>>>>>>> @@ -43,6 +43,24 @@ properties:
>>>>>>>> - enum:
>>>>>>>> - nvidia,tegra186-bpmp
>>>>>>>>
>>>>>>>> + interconnects:
>>>>>>>> + $ref: /schemas/types.yaml#/definitions/phandle-array
>>>>>>>> + description: A list of phandle and specifier pairs that describe the
>>>>>>>> + interconnect paths to and from the BPMP.
>>>>>>>> +
>>>>>>>> + interconnect-names:
>>>>>>>> + $ref: /schemas/types.yaml#/definitions/non-unique-string-array
>>>>>>>> + description: One string for each pair of phandle and specifier in the
>>>>>>>> + "interconnects" property.
>>>>>>>> + # XXX We need at least one of these to be named dma-mem so that the core
>>>>>>>> + # will set the DMA mask based on the DMA parent, but all of these go to
>>>>>>>> + # system memory eventually.
>>>>>>>> + items:
>>>>>>>> + - const: dma-mem
>>>>>>>> + - const: dma-mem
>>>>>>>> + - const: dma-mem
>>>>>>>> + - const: dma-mem
>>>>>
>>>>> Names should be unique, otherwise it's not possible to retrieve ICC path
>>>>> other than the first one.
>>>>
>>>> Yeah, I know, that's why there's an XXX comment. =) I just wasn't sure
>>>> what else to put there and thought this kinda made it clear that it was
>>>> only half-baked.
>>>>
>>>>>>>> iommus:
>>>>>>>> $ref: /schemas/types.yaml#/definitions/phandle-array
>>>>>>>> description: |
>>>>>>>> @@ -152,8 +170,43 @@ additionalProperties: false
>>>>>>>>
>>>>>>>> examples:
>>>>>>>> - |
>>>>>>>> + #include <dt-bindings/clock/tegra186-clock.h>
>>>>>>>> #include <dt-bindings/interrupt-controller/arm-gic.h>
>>>>>>>> #include <dt-bindings/mailbox/tegra186-hsp.h>
>>>>>>>> + #include <dt-bindings/memory/tegra186-mc.h>
>>>>>>>> +
>>>>>>>> + mc: memory-controller@2c00000 {
>>>>>>>> + compatible = "nvidia,tegra186-mc";
>>>>>>>> + reg = <0x02c00000 0xb0000>;
>>>>>>>> + interrupts = <GIC_SPI 223 IRQ_TYPE_LEVEL_HIGH>;
>>>>>>>> + status = "disabled";
>>>>>>>> +
>>>>>>>> + #interconnect-cells = <1>;
>>>>>>>> + #address-cells = <2>;
>>>>>>>> + #size-cells = <2>;
>>>>>>>> +
>>>>>>>> + ranges = <0x02c00000 0x0 0x02c00000 0x0 0xb0000>;
>>>>>>>> +
>>>>>>>> + /*
>>>>>>>> + * Memory clients have access to all 40 bits that the memory
>>>>>>>> + * controller can address.
>>>>>>>> + */
>>>>>>>> + dma-ranges = <0x0 0x0 0x0 0x100 0x0>;
>>>>>>>> +
>>>>>>>> + #memory-controller-cells = <0>;
>>>>>>>> +
>>>>>>>> + emc: external-memory-controller@2c60000 {
>>>>>>>> + compatible = "nvidia,tegra186-emc";
>>>>>>>> + reg = <0x0 0x02c60000 0x0 0x50000>;
>>>>>>>> + interrupts = <GIC_SPI 224 IRQ_TYPE_LEVEL_HIGH>;
>>>>>>>> + clocks = <&bpmp TEGRA186_CLK_EMC>;
>>>>>>>> + clock-names = "emc";
>>>>>>>> +
>>>>>>>> + #interconnect-cells = <0>;
>>>>>>>> +
>>>>>>>> + nvidia,bpmp = <&bpmp>;
>>>>>>>> + };
>>>>>>>> + };
>>>>>>>>
>>>>>>>> hsp_top0: hsp@3c00000 {
>>>>>>>> compatible = "nvidia,tegra186-hsp";
>>>>>>>> @@ -187,6 +240,12 @@ examples:
>>>>>>>>
>>>>>>>> bpmp {
>>>>>>>> compatible = "nvidia,tegra186-bpmp";
>>>>>>>> + interconnects = <&emc &mc TEGRA186_MEMORY_CLIENT_BPMPR>,
>>>>>>>> + <&mc TEGRA186_MEMORY_CLIENT_BPMPW &emc>,
>>>>>>>> + <&emc &mc TEGRA186_MEMORY_CLIENT_BPMPDMAR>,
>>>>>>>> + <&mc TEGRA186_MEMORY_CLIENT_BPMPDMAW &emc>;
>>>>>
>>>>> I don't think this is a correct definition of the ICC paths because the
>>>>> first node-MC_ID pair should define the source, second pair is the final
>>>>> destination. Then the interconnect core builds (by itself) the path from
>>>>> MC client to MC and finally to EMC based on the given source /
>>>>> destination. Please see my v1 patchset for the example.
>>>>
>>>> Okay, sounds like "source" in this case means the initiator of the
>>>> transaction and destination is the target of the transaction. I had
>>>> interpreted the "source" as the "source location" of the transaction (so
>>>> for reads the source would be the system memory via the EMC, and for
>>>> writes the source would be the memory client interface).
>>>
>>> Yes, exactly. Maybe it would be more correct to call these pairs
>>> initiator/target or master/slave.
>>
>> Do you want me to extend the bindings documentation to mention these
>> alternative names?
>>
>>>> Yeah, I think that makes sense. It was also pointed out to me (offline)
>>>> that the above doesn't work as intented for the use-case where I really
>>>> need it. The primary reason why I need these "dma-mem" interconnect
>>>> paths is so that the memory controller is specified as the "DMA parent"
>>>> for these devices, which is important so that the DMA masks can be
>>>> correctly set. Having the &emc reference in the first slot breaks that.
>>>> Your suggestion makes sense when interpreting the terminology
>>>> differently and it fixes the DMA parent issue (at least partially).
>>>>
>>>>> It should look somewhat like this:
>>>>>
>>>>> interconnects =
>>>>> <&mc TEGRA186_MEMORY_CLIENT_BPMPR &emc TEGRA_ICC_EMEM>,
>>>>> <&mc TEGRA186_MEMORY_CLIENT_BPMPW &emc TEGRA_ICC_EMEM>,
>>>>> <&mc TEGRA186_MEMORY_CLIENT_BPMPDMAR &emc TEGRA_ICC_EMEM>,
>>>>> <&mc TEGRA186_MEMORY_CLIENT_BPMPDMAW &emc TEGRA_ICC_EMEM>;
>>>>>
>>>>> interconnect-names = "bpmpr", "bpmpw", "bpmpdmar", "bpmpdmaw";
>>>
>>> This looks better to me.
>>>
>>>> I'm not sure if that TEGRA_ICC_EMEM makes a lot of sense. It's always
>>>> going to be the same and it's arbitrarily defined, so it's effectively
>>>> useless. But other than that it looks good.
>>>
>>> Well, in most cases the target would be the EMEM, so that's fine. I have seen
>>> that other vendors that may have an additional internal memory, especially
>>> dedicated to some DSPs and in such cases the bandwidth needs are different for
>>> the two paths (to internal memory and DDR).
>>
>> Most chips have a small internal memory that can be used, though it
>> seldomly is. However, in that case I would expect the target to be a
>> completely different device, so it'd look more like this:
>>
>> interconnects = <&mc TEGRA186_MEMORY_CLIENT_BPMPR &iram>,
>> ...;
>>
>> I don't think EMEM has any "downstream" other than external memory.
>
> The node ID should be mandatory in terms of interconnect, even if it's a
> single node. EMC (provider) != EMEM (endpoint).
>
Thinking a bit more about how to define the ICC, I'm now leaning to a
variant like this:
interconnects =
<&mc TEGRA186_MEMORY_CLIENT_BPMP &emc TEGRA_ICC_EMEM>,
<&mc TEGRA186_MEMORY_CLIENT_BPMPR>,
<&mc TEGRA186_MEMORY_CLIENT_BPMPW>,
<&mc TEGRA186_MEMORY_CLIENT_BPMPDMAR>,
<&mc TEGRA186_MEMORY_CLIENT_BPMPDMAW>;
interconnect-names = "dma-mem", "read", "write", "dma-read", "dma-write";
Looks like there is a problem with having a full MC-EMEM path being
defined for each memory client.. it's not very practical in terms of
memory frequency scaling.
Take Display Controller for example, it has a memory client for each
display (overlay) plane. If planes are not overlapping on the displayed
area, then the required total memory bandwidth equals to the peak
bandwidth selected among of the visible planes. But if planes are
overlapping, then the bandwidths of each overlapped plane are
accumulated because overlapping planes will issue read request
simultaneously for the overlapping areas.
The Memory Controller doesn't have any knowledge about the Display
Controller's specifics. Thus in the end it should be a responsibility of
Display Controller's driver to calculate the required bandwidth for the
hardware unit, since only the driver has all required knowledge about
planes overlapping state and whatnot.
The similar applies to multimedia things, like GPU or Video Decoder.
They have multiple memory clients and (I'm pretty sure that) nobody is
going to calculate memory bandwidth requirements for every client, it's
simply impractical.
So, I'm suggesting that we should have a single "dma-mem" ICC path for
every hardware unit.
The rest of the ICC paths could be memory_client -> memory_controller
paths, providing knobs for things like MC arbitration (latency)
configuration for each memory client. I think this variant of
description is actually closer to the hardware, since the client's
arbitration configuration ends in the Memory Controller.
Please let me know what do you think about the above, thanks in advance.
