On Tue, 11 Feb 2020 at 02:03, Stefano Stabellini via System-dt system-dt@lists.openampproject.org wrote:

Hi all,

During the last system device tree call, we agreed it would be helpful to have a document explaining how to use the proposed new device tree bindings to describe heterogeneous systems and solve typical problems, such as memory reservations for multiple domains, multiple interrupts controllers, etc.

Tomas and I wrote the following document (also attached: feel free to use pandoc to convert it into html if you prefer to read it that way). It includes an introduction to system device tree, a short description of the bindings, and how to use them to solve common problems. I hope that it will help create a common understanding of the problems we are trying to solve and the potential solutions. I also attached the full system device tree example as reference.

Cheers,

Stefano

System Device Tree Concepts

System Device Trees extends traditional Device Trees to handle heterogeneous SoCs with multiple CPUs and Execution Domains. An Execution Domain can be seen as an address space that is running a software image, whether an operating system, a hypervisor or firmware that has a set of cpus, memory and devices attached to it. I.e. Each individual CPU/core that is not part of an SMP cluster is a separate Execution Domain as is the different Execution Levels on an ARMv8-A architecture. Trusted and not trusted environment can also be viewed as separate Execution Domains.

A design goal of System Device Trees is that no current client of Device Trees should have to change at all, unless it wants to take advantage of the extra information. This means that Linux in particular does not need to change since it will see a Device Tree that it can handle with the current implementation, potentially with some extra information it can ignore.

System Device Trees must handle two types of heterogeneous additions:

1. Being able to specify different cpu clusters and the actual memory and devices hard-wired to them
   • This is done through the new Hardware Descriptions, such as "cpu,cluster" and "indirect-bus"
   • This information is provided by the SoC vendor and is typically fixed for a given SoC/board
2. Being able to assign hardware resources that can be configured by software to be used by one or more Execution Domains
   • This is done through the Execution Domain configuration
   • This information is provided by a System Architect and will be different for different use cases, even for the same board
     • E.g. How much memory and which devices goes to Linux vs. an RTOS can be different from one boot to another
   • This information should be separated from the hard-wired information for two reasons
     • A different persona will add and edit the information
     • Configuration should be separated from specification since it has a different rate of change

The System Device Trees and Execution Domain information are used in two major use cases:

1. Exclusively on the host by using a tool like Lopper that will "prune" the System Device Tree
   • Each domain will get its own "traditional" Device Tree that only sees one address space and has one "cpus" node, etc.
   • Lopper has pluggable backends to it can also generate information for clients that is using a different format
     • E.g. It can generate a bunch of "#defines" that can be included and compiled in to an RTOS
2. System Device Trees can be used by a "master" target environment that manages multiple Execution Domains:
   • a firmware that can set up hardware protection and use it to restart individual domains
     • E.g. Protect the Linux memory so the R5 OS can't reach it
   • any other operating system or hypervisor that has sub-domains
     • E.g. Xen can use the Execution Domains to get info about the Xen guests (also called domains)
     • E.g. Linux could use the default domain for its own configuration and the domains to manage other CPUs
     • Since System Device Trees are backwards compatible with Device Trees, the only changes needed in Linux would be any new code taking advantage of the Domain information
     • a default master has access to all resources (CPUs, memories, devices), it has to make sure it stops using the resource itself when it "gives it away" to a sub-domain

There is a concept of a default Execution Domain in System Device Trees, which corresponds to /cpus. The default domain is compatible with the current traditional Device Tree. It is useful for a couple of reasons:

1. As a way to specify the default place to assign added hardware (see use case #1)
   • A default domain does not have to list the all the HW resources allocated to it. It gets everything not allocated elsewhere by Lopper.
   • This minimizes the amount of information needed in the Domain configuration.
   • This is also useful for dynamic hardware such as add-on boards and FPGA images that are adding new devices.
2. The default domain can be used to specify what a master environment sees (see use case #2)
   • E.g. the default domain is what is configuring Linux or Xen, while the other domains specify domains to be managed by the master

System Device Tree Hardware Description

To turn system device tree into a reality we are introducing a few new concepts. They enable us to describe a system with multiple cpus clusters and potentially different address mappings for each of them (i.e. a device could be seen at different addresses from different cpus clusters).

The new concepts are:

• Multiple top level "cpus,cluster" nodes to describe heterogeneous CPU clusters.
• "indirect-bus": a new type of bus that does not automatically map to the parent address space (i.e. not automatically visible).
• An "address-map" property to express the different address mappings of the different cpus clusters and to map indirect-buses.

The following is a brief example to show how they can be used together:

    /* default cluster */
    cpus {
            cpu@0 {
            };
            cpu@1 {
            };
    };

    /* additional R5 cluster */
    cpus_r5: cpus-cluster@0 {
            compatible = "cpus,cluster";

            /* specifies address mappings */
            address-map = <0xf9000000 &amba_rpu 0xf9000000 0x10000>;

            cpu@0 {
            };

            cpu@1 {
            };
    };

    amba_rpu: indirect-bus@f9000000 {
            compatible = "indirect-bus";
    };

In this example we can see:

• two cpus clusters, one of them is the default top-level cpus node
• an indirect-bus "amba_rpu" which is not visible to the top-level cpus node
• the cpus_r5 cluster can see amba_rpu because it is explicitly mapped using the address-map property

Devices only physically accessible from one of the two clusters should be placed under an indirect-bus as appropriate. For instance, in the following example we can see how interrupts controllers are expressed:

    /* default cluster */
    cpus {
    };

    /* additional R5 cluster */
    cpus_r5: cpus-cluster@0 {
            compatible = "cpus,cluster";

            /* specifies address mappings */
            address-map = <0xf9000000 &amba_rpu 0xf9000000 0x10000>;
    };

    /* bus only accessible by cpus */
    amba_apu: bus@f9000000 {
            compatible = "simple-bus";

            gic_a72: interrupt-controller@f9000000 {
            };
    };

    /* bus only accessible by cpus_r5 */
    amba_rpu: indirect-bus@f9000000 {
            compatible = "indirect-bus";

            gic_r5: interrupt-controller@f9000000 {
            };
    };

gic_a72 is accessible by /cpus, but not by cpus_r5, because amba_apu is not present in the address-map of cpus_r5.

gic_r5 is visible to cpus_r5, because it is present in the address map of cpus_r5. gic_r5 is not visible to /cpus because indirect-bus doesn't automatically map to the parent address space, and /cpus doesn't have an address-map property in the example.

Relying on the fact that each interrupt controller is correctly visible to the right cpus cluster, it is possible to express interrupt routing from a device to multiple clusters. For instance:

    amba: bus@f1000000 {
            compatible = "simple-bus";
            ranges;

            #interrupt-cells = <3>;
            interrupt-map-pass-thru = <0xffffffff 0xffffffff 0xffffffff>;
            interrupt-map-mask = <0x0 0x0 0x0>;
            interrupt-map = <0x0 0x0 0x0 &gic_a72 0x0 0x0 0x0>,
                            <0x0 0x0 0x0 &gic_r5 0x0 0x0 0x0>;

            can@ff060000 {
                    compatible = "xlnx,canfd-2.0";
                    reg = <0x0 0xff060000 0x0 0x6000>;
                    interrupts = <0x0 0x14 0x1>;
            };
    };

In this example, all devices under amba, including can@ff060000, have their interrupts routed to both gic_r5 and gic_a72.

Memory only physically accessible by one of the clusters can be placed under an indirect-bus like any other device types. However, normal memory is usually physically accessible by all clusters. It is just a software configuration that splits memory into ranges and assigns a range for each execution domain. Software configurations are explained below.

Execution Domains

An execution domain is a collection of software, firmware, and board configurations that enable an operating system or an application to run a cpus cluster. With multiple cpus clusters in a system it is natural to have multiple execution domains, at least one per cpus cluster. There can be more than one execution domain for each cluster, with virtualization or non-lockstep execution (for cpus clusters that support it). Execution domains are configured and added at a later stage by a software architect.

Execution domains are expressed by a new node "openamp,domain" compatible. Being a configuration rather than a description, their natural place is under /chosen or under a similar new top-level node. In this example, I used /domains:

    domains {
            openamp_r5 {
                    compatible = "openamp,domain-v1";
                    cpus = <&cpus_r5 0x2 0x80000000>;
                    memory = <0x0 0x0 0x0 0x8000000>;
                    access = <&can@ff060000>;
            };
    };

An openamp,domain node contains information about:

• cpus: physical cpus on which the software is running on
• memory: memory assigned to the domain
• access: any devices configured to be only accessible by a domain

The access list is an array of links to devices that are configured to be only accessible by an execution domain, using bus firewalls or similar technologies.

The memory range assigned to an execution domain is expressed by the memory property. It needs to be a subset of the physical memory in the system. The memory property can also be used to express memory sharing between domains:

    domains {
            openamp_r5 {
                    compatible = "openamp,domain-v1";
                    memory = <0x0 0x0 0x0 0x8000000 0x8 0x0 0x0 0x10000>;
            };
            openamp_a72 {
                    compatible = "openamp,domain-v1";
                    memory = <0x0 0x8000000 0x0 0x80000000 0x8 0x0 0x0 0x10000>;
            };
    };

Are those domains also define a "reset domain"? I mean, you may want to selectively reset a domain (within a particular clock domain). Or shall reset-domains be defined and then domains would be defined on subsets of those?

In this example, a 16 pages range starting at 0x800000000 is shared between two domains.

In a system device tree without a default cpus cluster (no top-level cpus node), lopper figures out memory assignment for each domain by looking at the memory property under each "openamp,domain" node. In a device tree with a top-level cpus cluster, and potentially a legacy OS running on it, we might want to "hide" the memory reservation for other clusters from /cpus. We can do that with /reserved-memory:

    reserved-memory {
            #address-cells = <0x2>;
            #size-cells = <0x2>;
            ranges;

            memory_r5@0 {
                    compatible = "openamp,domain-memory-v1";
                    reg = <0x0 0x0 0x0 0x8000000>;
            };
    };

The purpose of memory_r5@0 is to let the default execution domain know that it shouldn't use the 0x0-0x8000000 memory range because it is reserved for use by other domains.

/reserved-memory and /chosen are top-level nodes dedicated to configurations, rather than hardware description. Each execution domain might need similar configurations, hence, chosen and reserved-memory are also specified under each openamp,domain node for domains specific configurations. The top-level /reserved-memory and /chosen nodes remain in place for the default execution domain. As an example:

    /chosen -> configuration for a legacy OS running on /cpus
    /reserved-memory -> reserved memory for a legacy OS running on /cpus

    /domains/openamp_r5/chosen -> configuration for the domain "openamp_r5"
    /domains/openamp_r5/reserved-memory -> reserved memory for "openamp_r5"--

System-dt mailing list System-dt@lists.openampproject.org https://lists.openampproject.org/mailman/listinfo/system-dt