Mailbox NS Agent Design Update

Organization

Arm Limited

Contact

tf-m@lists.trustedfirmware.org

Background

The SPE component that maintains the non-secure clients’ request is called ‘NS Agent’ in TF-M. Besides the Trustzone-based isolation mechanism, there is one other isolation mechanism that implements individual PEs in physically isolated cores respectively. NSPE and SPE transfer non-secure client requests via inter-processor communication based on mailboxes. The component that handles inter-processor communication messages is called Mailbox NS Agent.

Note

There may be hardware components and software solutions containing ‘mailbox’ in their names. The concept mailbox in this document represent the mechanism described above, which is not referring to the external concepts.

When the first version Mailbox NS Agent was introduced, the generic FFM interrupt handling was not ready. Hence a customized solution Multiple Core is implemented. This customized implementation:

  • Perform customized operations on SPM internal data in a deferred interrupt handler.

  • Process mailbox operations as special cases in SPM common logic.

These behaviours couple SPM tightly with mailbox logic, which bring issues for maintenance. To address the issue, an updated design shall:

  • Make SPM manage other components in a unified way (For example, it is simpler for SPM if all non-SPM components under the IPC model act as processes.)

  • Can use FF-M compliant interrupt mechanism and APIs.

Following the above guidelines makes the Mailbox NS Agent work like a partition. The agent has an endless loop and waits for signals, calls FFM API based on the parsing result on the communication messages. But there are still issues after looking closer to the requirements of the agent:

  • SPM treats FFM Client API caller’s ID as the client ID. While the mailbox NS agent may represent multiple non-secure clients. Hence it needs to tell SPM which non-secure client it is representing, and the default FFM Client API does not have such capability.

  • FFM Client API blocks caller before the call is replied; while the mailbox NS Agent needs to respond to the non-secure interrupts in time. Blocking while waiting for a reply may cause the non-secure communication message not to be handled in time.

Extra design items need to be added to address the issues.

Design Update

The below figure shows the overall design to cover various component types. NS Agents are the implementation-defined components that provide FF-M compliant Client API to the non-secure clients. Hence from the view of the non-secure clients, the FF-M client API behaviour follows the FF-M definition. And NS Agent needs customization in SPM since it has extra requirements compared to a generic secure partition.

../../_images/mailbox_ns_agent1.svg

Figure 4: Component types and the callable API set

Note

3 non-SPM component types here: FFM-compliant Secure Partition (aka partition), Trustzone-based NS Agent (aka Trustzone NS Agent) and mailbox-based NS Agent (aka Mailbox NS Agent). Trustzone NS Agent is mentioned here for the comparison purpose. The implementation details for this NS agent type is not introduced here.

To make the programming model close to the FFM compliance, the Mailbox NS Agent is designed as:

  • Working like a standard Secure Partition under the IPC model, has one single thread, can call FFM standard API.

  • Having a manifest file to describe the attributes and resources and a positive valued Partition ID in the manifest.

Services rely on the client_id to apply policy-checking, hence SPM needs to know which client_id the mailbox NS Agent is representing when mailbox API is calling Client API. The standard API treats the caller as the client of the service, which means that a specific API is required to support identifying the represented non-secure client. SPM sets the non-secure client_id into the message right at the moment the message is going to be sent. Before this point, SPM performs the call based on the agent’s ID.

This specifc Agent API is non-blocking, unlike the standard FF-M Client APIs. This can improve the communication efficiency between NS clients and mailbox NS agents. With this mechanism, extra signals and APIs for message acknowledges are also required.

Note

A standard Secure Partition gets errors when calling the Agent API.

Updated programming interfaces

These Client APIs are expanded from the standard Client APIs:

  • agent_psa_connect() is extended from psa_connect().

  • agent_psa_close() is extended from psa_close().

  • agent_psa_call() is extended from psa_call().

And to cooperate with the changed behaviour of these APIs, extra defined signals and types are also involved.

Note

Namespace agent is involved for NS Agent callable API; namespace tfm is involved for TF-M specific concepts. Even though agent is TF-M specific at the current stage, it is proposed to be a common concept for general FF-M compliant implementations, hence assigning agent for proposed API and data structures.

Input and output vectors

When non-secure clients call psa_call(), a mailbox message containing psa_call() parameters is delivered into Mailbox NS Agent and the agent needs to extract parameters from the message and then call agent_psa_call(). Revisit the psa_call() prototype to see the parameters:

psa_status_t psa_call(psa_handle_t handle,
                      int32_t type,
                      const psa_invec *in_vec,
                      size_t in_len,
                      psa_outvec *out_vec,
                      size_t out_len);

Interface agent_psa_call() has 4 arguments only, to avoid ABI complexity when more than 4 arguments get involved. Then input and output vectors must be squashed into a new type to achieve this squashing. There are several scenarios to be considered:

  • In the shared-memory-based mailbox scheme, the non-secure interface layer puts psa_invec and psa_outvec instances and their members in the shared memory. The pointers of these items are delivered to Mailbox NS Agent for the agent’s referencing. All vectors are non-secure in this case.

  • Still, in the shared-memory-based mailbox scheme, but this time not all parameters are prepared by non-secure clients. Some of the items are allocated by the agent. Secure and non-secure vectors get mixed in this case.

  • In the scheme that a memory address can not be delivered (A serial interface e.g.), Mailbox NS Agent allocates psa_invec and psa_outvec in local memory to collect pointers and sizes of received buffers. All the vectors are secure in this case.

Based on these scenarios, and the case that an agent might access depending services with the agent itself’s identifier and secure vectors, a straight conclusion is that the memory source information - secure or non-secure - for input and output vectors and their pointing memories need to be indicated by the agent so that SPM can perform a memory check towards given vectors.

Most of the SPE platforms have the capability to identify if a given memory pointer is secure or not, which makes this indication look duplicated. But the indication is necessary for these scenarios:

  • The SPE platform identifies the memory source by the address mapped into SPE’s address space, but this mapping happens inside the agent instead of SPM. It is the agent who receives the mailbox data and maps it, so it knows which addresses need mapping and the others do not.

  • The SPE platform just does not have such capability. The addresses from the mailbox can be treated as non-secure always but there are cases that the agent itself needs to access services with its own memory instead of representing the non-secure clients.

To cover the above mentioned scenarios, guidelines are listed for input and output vector processing:

  • The agent needs to tell SPM where vectors and descriptors come from, to assist SPM performs a proper memory checking. The source information is encoded in the parameter control.

  • When SPE platforms have the capability to identify the memory sources, platforms can decide whether to skip the indication or not, in the HAL.

A composition type is created for packing the vectors:

struct client_param_t {
  int32_t         ns_client_id_stateless;
  const psa_invec *p_invecs;
  psa_outvec      *p_outvecs;
};

ns_client_id_stateless indicates the non-secure client id when the client is accessing a stateless service. This member is ignored if the target service is a connection-based one.

Note

The vectors and non-secure client ID are recorded in the internal handle. Hence it is safe to claim client_param_t instance as local variable.

Agent-specific Client API

agent_psa_connect() and agent_psa_close() are the APIs added to support agent forwarding NS requests.

psa_handle_t agent_psa_connect(uint32_t sid, uint32_t version,
                               int32_t ns_client_id, const void *client_data);
psa_status_t agent_psa_close(psa_handle_t handle, int32_t ns_client_id);

One extra parameter ns_client_id added to tell SPM which NS client the agent is representing when API gets called. It is recorded in the handle association data in SPM and requires to be a negative value; ZERO or positive values are invalid non-secure client IDs, SPM does not use these invalid IDs in the message. Instead, it puts the agent’s ID into the messaging in this case. This mechanism can provide chances for the agents calling APIs for their own service accessing and API works asynchronously.

As mentioned, the standard FFM Client service accessing API are blocked until the IPC message gets replied to. While this API returns immediately without waiting for acknowledgement. Unless an error occurred, these agent-specific API returns PSA_SUCCESS always. The replies for these access requests are always fetched initiative by the agent with a psa_get().

psa_status_t agent_psa_call(psa_handle_t handle, uint32_t control,
                            const struct client_param_t   *params,
                            const void   *client_data_stateless);

Compared to the standard psa_call(), this API:

  • Squashes the psa_invec, psa_outvec, and ns_client_id_stateless into parameter params.

  • One extra parameter client_data_stateless for agent_psa_call() stands for the auxiliary data added. This member is ignored for connection-based services because agent_psa_connect() already assigned one in the connected handle.

  • Has a composited agrument control.

The encoding scheme for control:

Bit(s)

Name

Description

27

NSIV

1: Input vectors in non-secure memory. 0: Secure memory.

24-26

IVNUM

Number of input vectors.

19

NSOV

1: Output vectors in non-secure memory. 0: Secure memory.

16-18

OVNUM

Number of output vectors.

0-15

type

signed 16-bit service type.

Note

control is a 32-bit unsigned integer and bits not mentioned in the table are reserved for future usage.

Agent-specific signal

To cooperate with the agent-specific API, one extra acknowledgement signal is defined:

#define ASYNC_MSG_REPLY            (0x00000004u)

This signal can be sent to agent type component only. An agent can call psa_get() with this signal to get one acknowledged message. This signal is cleared when all queued messages for the agent have been retrieved using psa_get(). SPM assembles the information into agent provided message object. For the stateless handle, the internal handle object is freed after this psa_get() call. The agent can know what kind of message is acknowledged by the type member in the psa_msg_t, and the client_data passed in is put in member rhandle. If no ‘ASYNC_MSG_REPLY’ signals pending, calling psa_get() gets panic.

Code Example

/*
 * The actual implementation can change this __customized_t freely, or
 * discard this type and apply some in-house mechanism - the example
 * here is to introduce how an agent works only.
 */
struct __customized_t {
    int32_t      type;
    int32_t      client_id;
    psa_handle_t handle;
    psa_handle_t status;
};

void mailbox_main(void)
{
    psa_signal_t   signals;
    psa_status_t   status;
    psa_msg_t      msg;
    struct client_param_t client_param;
    struct __customized_t ns_msg;

    while (1) {
        signals = psa_wait(ALL, BLOCK);

        if (signals & MAILBOX_INTERRUPT_SIGNAL) {
            /* NS memory check needs to be performed. */
            __customized_platform_get_mail(&ns_msg);

            /*
             * MACRO 'SID', 'VER', 'NSID', 'INVEC_LEN', 'OUTVEC_LEN', and
             * 'VECTORS' represent necessary information extraction from
             * 'ns_msg', put MACRO names here and leave the details to the
             * implementation.
             */
            if (ns_msg.type == PSA_IPC_CONNECT) {
                status = agent_psa_connect(SID(ns_msg), VER(ns_msg),
                                           NSID(ns_msg), &ns_msg);
            } else if (ns_msg.type == PSA_IPC_CLOSE) {
                status = agent_psa_close(ns_msg.handle, NSID(ns_msg));
            } else {
                /* Other types as call type and let API check errors. */
                client_param.ns_client_id_stateless = NSID(ns_msg);

                /*
                 * Use MACRO to demonstrate two cases: local vector
                 * descriptor and direct descriptor forwarding.
                 */

                /* Point to vector pointers in ns_msg. */
                PACK_VECTOR_POINTERS(client_param, ns_msg);
                status = agent_psa_call(ns_msg.handle,
                                        PARAM_PACK(ns_msg.type,
                                                   INVEC_LEN(ns_msg),
                                                   OUTVEC_LEN(ns_msg))|
                                                   NSIV | NSOV,
                                        &client_param,
                                        &ns_msg);
            }
            /*
             * The service access reply is always fetched by a later
             * `psa_get` hence here only errors need to be dispatched.
             */
            error_dispatch(status);

        } else if (signals & ASYNC_MSG_REPLY) {
            /* The handle is freed for stateless service after 'psa_get'. */
            status        = psa_get(ASYNC_MSG_REPLY, &msg);
            ms_msg        = msg.rhandle;
            ns_msg.status = status;
            __customized_platform__send_mail(&ns_msg);
        }
    }
}

Note

__customized* API are implementation-specific APIs to be implemented by the mailbox Agent developer.

Customized manifest attribute

Two extra customized manifest attributes are added:

Name

Description

ns_agent

Indicate if manifest owner is an Agent.

ns_client_ids

Possible non-secure Client ID values (<0).

Attribute ‘ns_client_ids’ can be a set of numbers, or it can use a range expression such as [min, max]. The tooling can detect ID overlap between multiple non-secure agents.

Note

Per-non-secure-client dependencies scheme is now assumed to be implemented by agent customization. Feedback if there are requirements for a unified scheme.

Manifest tooling update

The manifest for agents involves specific keys (‘ns_agent’ e.g.), these keys give hints about how to achieve out-of-FFM partitions which might be abused easily by developers, for example, claim partitions as agents. Some restrictions need to be applied in the manifest tool to limit the general secure service development referencing these keys.

Note

The limitations can mitigate the abuse but can’t prevent it, as developers own all the source code they are working with.

One mechanism: adding a confirmation in the partition list file.

"description": "Non-Secure Mailbox Agent",
"manifest": "${CMAKE_SOURCE_DIR}/secure_fw/partitions/ns_agent_mailbox/ns_agent_mailbox.yaml",
"non_ffm_attributes": "ns_agent", "other_option",

non_ffm_attributes tells the manifest tool that ns_agent is valid in ns_agent_mailbox.ymal. Otherwise, the manifest tool reports an error when a non-agent service abuses ns_agent in its manifest.

Runtime programming characteristics

Mailbox agent shall not be blocked by Agent-specific APIs. It can be blocked when:

  • It is calling standard PSA Client APIs.

  • It is calling psa_wait().

IDLE processing

Only ONE place is recommended to enter IDLE. The place is decided based on the system topologies:

  • If there is one Trustzone-based NSPE, this NSPE is the recommended place no matter how many mailbox agents are in the system.

  • If there are only mailbox-based NSPEs, entering IDLE can happen in one of the mailbox agents.

The solution is:

  • An IDLE entering API is provided in SPRTL.

  • A partition without specific flag can’t call this API.

  • The manifest tooling counts the partitions with this specific flag, and assert errors when multiple instances are found.


Copyright (c) 2022-2024, Arm Limited. All rights reserved. Copyright (c) 2023 Cypress Semiconductor Corporation (an Infineon company) or an affiliate of Cypress Semiconductor Corporation. All rights reserved.