Re: [PATCH v12 0/7] Introduce on-chip interconnect API

Hi Olof,

On 9.12.18 2:33, Olof Johansson wrote:

Hi Georgi,

On Sat, Dec 8, 2018 at 9:02 AM Georgi Djakov [off-list ref] wrote:

quoted

Modern SoCs have multiple processors and various dedicated cores (video, gpu,
graphics, modem). These cores are talking to each other and can generate a
lot of data flowing through the on-chip interconnects. These interconnect
buses could form different topologies such as crossbar, point to point buses,
hierarchical buses or use the network-on-chip concept.

These buses have been sized usually to handle use cases with high data
throughput but it is not necessary all the time and consume a lot of power.
Furthermore, the priority between masters can vary depending on the running
use case like video playback or CPU intensive tasks.

Having an API to control the requirement of the system in terms of bandwidth
and QoS, so we can adapt the interconnect configuration to match those by
scaling the frequencies, setting link priority and tuning QoS parameters.
This configuration can be a static, one-time operation done at boot for some
platforms or a dynamic set of operations that happen at run-time.

This patchset introduce a new API to get the requirement and configure the
interconnect buses across the entire chipset to fit with the current demand.
The API is NOT for changing the performance of the endpoint devices, but only
the interconnect path in between them.

The API is using a consumer/provider-based model, where the providers are
the interconnect buses and the consumers could be various drivers.
The consumers request interconnect resources (path) to an endpoint and set
the desired constraints on this data flow path. The provider(s) receive
requests from consumers and aggregate these requests for all master-slave
pairs on that path. Then the providers configure each participating in the
topology node according to the requested data flow path, physical links and
constraints. The topology could be complicated and multi-tiered and is SoC
specific.

This patch series description fails to describe why you need a brand
new subsystem for this instead of either using one of the current
ones, or adapting it to fit the needs you have.

Primarily, I'm wondering what's missing from drivers/devfreq to fit your needs?

The devfreq subsystem seems to be more oriented towards a device (like GPU or CPU) that controls the power/performance characteristics by itself and not the performance of other devices. The main problem of using it is that it's using a reactive approach - for example monitor some performance counters and then reconfigure bandwidth after some bottleneck has already occurred. This is suboptimal and might not work well. The new solution does the opposite by allowing drivers to express their needs in advance and be proactive. Devfreq also does not seem suitable for configuring complex, multi-tiered bus topologies and aggregating constraints provided by drivers.

The series also doesn't seem to provide any kind of indication how
this will be used by end points. You have one driver for one SoC that
just contains large tables that are parsed at probe time, but no
driver hooks anywhere that will actually change any settings depending
on use cases. Also, the bindings as posted don't seem to include any
of this kind of information. So it's hard to get a picture of how this
is going to be used in reality, which makes it hard to judge whether
it is a good solution or not.

Here are links to some of the examples that are on the mailing list already. I really should have  included them in the cover letter. 
https://lkml.org/lkml/2018/12/7/584
https://lkml.org/lkml/2018/10/11/499
https://lkml.org/lkml/2018/9/20/986
https://lkml.org/lkml/2018/11/22/772

Platforms drivers for different SoCs are available:
https://lkml.org/lkml/2018/11/17/368
https://lkml.org/lkml/2018/8/10/380
There is a discussion on linux-pm about supporting also Tegra platforms in addition to NXP and Qualcomm.

Overall, exposing all of this to software is obviously a nightmare
from a complexity point of view, and one in which it will surely be
very very hard to make the system behave properly for generic
workloads beyond benchmark tuning.

It allows the consumer drivers to dynamically express their performance needs in the system in a more fine grained way (if they want/need to) and this helps the system to keep the lowest power profile. This has already been done for a long time in various different kernels shipping with Android devices, for example, and basically every vendor uses a different custom approach. So I believe that this is doing the generalization that was needed.

Having more information about the above would definitely help tell if
this whole effort is a step in the right direction, or if it is
needless complexity that is better solved in other ways.

Sure, hope that this answers your questions.

Thanks,
Georgi

-Olof