                          The Linux RapidIO Subsystem

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The RapidIO standard is a packet-based fabric interconnect standard designed for
use in embedded systems. Development of the RapidIO standard is directed by the
RapidIO Trade Association (RTA). The current version of the RapidIO specification
is publicly available for download from the RTA web-site [1].

This document describes the basics of the Linux RapidIO subsystem and provides
information on its major components.

1 Overview
----------

Because the RapidIO subsystem follows the Linux device model it is integrated
into the kernel similarly to other buses by defining RapidIO-specific device and
bus types and registering them within the device model.

The Linux RapidIO subsystem is architecture independent and therefore defines
architecture-specific interfaces that provide support for common RapidIO
subsystem operations.

2. Core Components
------------------

A typical RapidIO network is a combination of endpoints and switches.
Each of these components is represented in the subsystem by an associated data
structure. The core logical components of the RapidIO subsystem are defined
in include/linux/rio.h file.

2.1 Master Port

A master port (or mport) is a RapidIO interface controller that is local to the
processor executing the Linux code. A master port generates and receives RapidIO
packets (transactions). In the RapidIO subsystem each master port is represented
by a rio_mport data structure. This structure contains master port specific
resources such as mailboxes and doorbells. The rio_mport also includes a unique
host device ID that is valid when a master port is configured as an enumerating
host.

RapidIO master ports are serviced by subsystem specific mport device drivers
that provide functionality defined for this subsystem. To provide a hardware
independent interface for RapidIO subsystem operations, rio_mport structure
includes rio_ops data structure which contains pointers to hardware specific
implementations of RapidIO functions.

2.2 Device

A RapidIO device is any endpoint (other than mport) or switch in the network.
All devices are presented in the RapidIO subsystem by corresponding rio_dev data
structure. Devices form one global device list and per-network device lists
(depending on number of available mports and networks).

2.3 Switch

A RapidIO switch is a special class of device that routes packets between its
ports towards their final destination. The packet destination port within a
switch is defined by an internal routing table. A switch is presented in the
RapidIO subsystem by rio_dev data structure expanded by additional rio_switch
data structure, which contains switch specific information such as copy of the
routing table and pointers to switch specific functions.

The RapidIO subsystem defines the format and initialization method for subsystem
specific switch drivers that are designed to provide hardware-specific
implementation of common switch management routines.

2.4 Network

A RapidIO network is a combination of interconnected endpoint and switch devices.
Each RapidIO network known to the system is represented by corresponding rio_net
data structure. This structure includes lists of all devices and local master
ports that form the same network. It also contains a pointer to the default
master port that is used to communicate with devices within the network.

3. Subsystem Initialization
---------------------------

In order to initialize the RapidIO subsystem, a platform must initialize and
register at least one master port within the RapidIO network. To register mport
within the subsystem controller driver initialization code calls function
rio_register_mport() for each available master port. After all active master
ports are registered with a RapidIO subsystem, the rio_init_mports() routine
is called to perform enumeration and discovery.

In the current PowerPC-based implementation a subsys_initcall() is specified to
perform controller initialization and mport registration. At the end it directly
calls rio_init_mports() to execute RapidIO enumeration and discovery.

4. Enumeration and Discovery
----------------------------

When rio_init_mports() is called it scans a list of registered master ports and
calls an enumeration or discovery routine depending on the configured role of a
master port: host or agent.

Enumeration is performed by a master port if it is configured as a host port by
assigning a host device ID greater than or equal to zero. A host device ID is
assigned to a master port through the kernel command line parameter "riohdid=",
or can be configured in a platform-specific manner. If the host device ID for
a specific master port is set to -1, the discovery process will be performed
for it.

The enumeration and discovery routines use RapidIO maintenance transactions
to access the configuration space of devices.

The enumeration process is implemented according to the enumeration algorithm
outlined in the RapidIO Interconnect Specification: Annex I [1].

The enumeration process traverses the network using a recursive depth-first
algorithm. When a new device is found, the enumerator takes ownership of that
device by writing into the Host Device ID Lock CSR. It does this to ensure that
the enumerator has exclusive right to enumerate the device. If device ownership
is successfully acquired, the enumerator allocates a new rio_dev structure and
initializes it according to device capabilities.

If the device is an endpoint, a unique device ID is assigned to it and its value
is written into the device's Base Device ID CSR.

If the device is a switch, the enumerator allocates an additional rio_switch
structure to store switch specific information. Then the switch's vendor ID and
device ID are queried against a table of known RapidIO switches. Each switch
table entry contains a pointer to a switch-specific initialization routine that
initializes pointers to the rest of switch specific operations, and performs
hardware initialization if necessary. A RapidIO switch does not have a unique
device ID; it relies on hopcount and routing for device ID of an attached
endpoint if access to its configuration registers is required. If a switch (or
chain of switches) does not have any endpoint (except enumerator) attached to
it, a fake device ID will be assigned to configure a route to that switch.
In the case of a chain of switches without endpoint, one fake device ID is used
to configure a route through the entire chain and switches are differentiated by
their hopcount value.

For both endpoints and switches the enumerator writes a unique component tag
into device's Component Tag CSR. That unique value is used by the error
management notification mechanism to identify a device that is reporting an
error management event.

Enumeration beyond a switch is completed by iterating over each active egress
port of that switch. For each active link, a route to a default device ID
(0xFF for 8-bit systems and 0xFFFF for 16-bit systems) is temporarily written
into the routing table. The algorithm recurs by calling itself with hopcount + 1
and the default device ID in order to access the device on the active port.

After the host has completed enumeration of the entire network it releases
devices by clearing device ID locks (calls rio_clear_locks()). For each endpoint
in the system, it sets the Discovered bit in the Port General Control CSR
to indicate that enumeration is completed and agents are allowed to execute
passive discovery of the network.

The discovery process is performed by agents and is similar to the enumeration
process that is described above. However, the discovery process is performed
without changes to the existing routing because agents only gather information
about RapidIO network structure and are building an internal map of discovered
devices. This way each Linux-based component of the RapidIO subsystem has
a complete view of the network. The discovery process can be performed
simultaneously by several agents. After initializing its RapidIO master port
each agent waits for enumeration completion by the host for the configured wait
time period. If this wait time period expires before enumeration is completed,
an agent skips RapidIO discovery and continues with remaining kernel
initialization.

5. References
-------------

[1] RapidIO Trade Association. RapidIO Interconnect Specifications.
    http://www.rapidio.org.
[2] Rapidio TA. Technology Comparisons.
    http://www.rapidio.org/education/technology_comparisons/
[3] RapidIO support for Linux.
    http://lwn.net/Articles/139118/
[4] Matt Porter. RapidIO for Linux. Ottawa Linux Symposium, 2005
    http://www.kernel.org/doc/ols/2005/ols2005v2-pages-43-56.pdf