Open SoC Debug Primer¶
About Open SoC Debug¶
Systems-on-Chip (SoCs) have become embedded deeply into our lives. Most of the time we enjoy the way they serve their purpose without getting in the way. Until they don’t. In those moments, we as engineers are reminded of the complex interplay between software and hardware in SoCs. We might pose questions like “How does my software execute on the chip?” or “Why is it showing this exact behavior?” To answer these questions we need insight into the system that executes the software. We gain this insight through the debug infrastructure integrated into the SoC.
Even though debug infrastructure is an essential part of any SoC design, most people consider creating it more of a necessary chore than an exciting endeavor. Therefore, most vendors today include debug infrastructure that follows one of two major specifications: ARM CoreSight and NEXUS 5001 (officially called “IEEE-ISTO 5001”). Unfortunately, none of these specifications are fully open, they cannot be used without any money involved.
The Open SoC Debug (OSD) specification was created to close this gap. Three key messages guide its design.
- OSD is a truly open (source) specification. Without any committee
membership required or royalty fees to be payed, anyone can freely
- share and modify the specification itself, and
- create and distribute implementations of the specification for any purpose.
- OSD is for debugging and tracing. A debugging infrastructure by itself is not a solution, but a toolbox providing the right tool for the task. Some bugs are best hunted using run-control debugging, some are better found using tracing. OSD supports both, enabling hardware and software developers to pick what’s best for their needs.
- OSD provides the common and enables the special. SoCs came to live because they allow reuse of components and specialization at the same time, letting hardware designers focus on the unique challenges without re-inventing the wheel. OSD follows this lead. Common IP blocks, interfaces and software tools can be re-used, and multiple extension vectors allow for easy customization where necessary.
By implementing Open SoC Debug, a SoC gains the following features (to a varying and implementation-defined degree).
- Support for run-control debugging, i.e. setting breakpoints and watchpoints and reading register values. In short, all you need to attach a debugger like GDB to the SoC.
- Support for tracing, i.e. non-intrusively observing the program execution on the SoC.
- Support for remotely controlling the SoC during development, e.g. starting the CPUs, resetting the system, and reading and writing the memories.
To provide these features, this specification defines
- an extensible debug system architecture, covering both hardware and host software,
- templates with well-defined interfaces for debug and trace IP blocks (“debug modules”),
- a set of common debug modules for the most frequent run-control debug and tracing tasks,
- a host-side software programming interface (API) for debug tools to interact with an OSD-enabled debug system.
In addition, implementations of many components described in the OSD specification are made available under a permissive open source license which can be used directly in custom designs.
OSD is an evolving effort. Currently, we target the first release of the base specification and module specification, that contain the following parts:
- Basic interfaces and transport protocols
- A generic and mandatory memory map for all debug modules to allow enumeration, capabilities and versioning
- Basic modules for run-control and trace debugging
This is just the start that covers the very basic functionality, but more features are planned to be added to the specification in the near future: tracing to memory instead of host, device traces, module triggering, cross-triggers, on-chip aggregation and filtering, sophisticated interconnects, just to mention a few.
About This Document¶
This document gives an overview of Open SoC Debug. The goal is to provide interested designers SoC hardware components as well as developers of debugging software tools a good understanding of the overall picture and the reasoning behind the design of OSD. This document is not the specification itself. Please refer to the individual sub-documents for the exact wording of the specification.
By implementing OSD, a SoC can easily be enhanced with advanced debug functionality. This section describes these features in more detail.
Run-control debugging, a.k.a. breakpoint debugging, “stop-and-stare” debugging, or just “debugging,” is the most common way of finding problems in software at early stages of development. Using software tools like the GNU Debugger (gdb) breakpoints can be set in the software code. If this point in the program reached, the program execution is stalled and the program control is handed over to the debugger. Using the debugger, a developer can now read register or memory values, print stack traces, and much more. To be efficient, run-control debugging functionality needs hardware support to stop the program execution at a given time. In addition, run-control debugging on SoC platforms is usually done remotely, i.e. the system is controlled from a host PC, as opposed to running the debugger directly on the SoC.
OSD contains all parts to add run-control debug support to a SoC. On the hardware side, OSD interfaces with the CPU core(s) to control its behavior. On the host side, OSD provides a daemon that GDB can connect to. The actual debugging is then handled by GDB and the usage of OSD is transparent to the software developer.
Today’s heterogeneous multi-core designs present new challenges to software developers. Concurrent software distributed across multiple CPUs and hardware accelerators, interacting with complex I/O interfaces and strict real-time requirements is the new normal. This results in new classes of bugs which are hard to find, like race conditions, deadlocks, and severely degraded performance for no obvious reason. To find such bugs, run-control debugging is not applicable: setting a breakpoint disturbs the temporal relationship between the different threads of execution. This disturbance to the program execution is called “probe effect” and can cause the original problem to disappear when searching for it, a phenomenon known as “Heisenbug.”
Tracing avoids these problems by unobtrusively monitoring the program execution and transferring the observations off-chip. There, the program flow can be reconstructed and the program behavior analyzed.
OSD comes with components to enable tracing for not only CPU cores, but also for any component in the SoC, such as memories, hardware accelerators, and interconnects.
Reading and writing memories is an essential tool during bring-up and debugging of a SoC. A typical use case is to write software to a program memory from the host PC, to avoid writing it for example to a SD card or flash memory and then resetting the system.
OSD ships with a module that can be attached to a memory to support reads and writes from and to memories.
Users of today’s debug systems know the pain: setting up a debugger on a host PC to communicate with the hardware often requires obscure configuration settings, secret switches and a bit of magic sauce to make it all work.
OSD is designed to be plug-and-play. All hardware components are self-describing. When a host connects to the system, it first enumerates all available components, and reads necessary configuration bits.
Timestamps are monotonically increasing numbers which are attached to events generated by the debug system. (They usually do not correspond in any way to wall-clock time.) Timestamps enable correlation of events in different parts of the chip with each other. Additionally, they can be used to restore order to events which are (for some reason) out of order when they arrive.
While timestamps are useful in many cases, adding them to all events generated by the debug system can significantly increase the overhead of such events.
Currently OSD supports timestamps which are full numbers of configurable width. Some debug modules can be configured to enable or disable timestamp generation.
The timestamping method used in OSD is referred to as “source timestamps” in some debug systems. Timestamps are added to the trace data at the source, as opposed to (e.g.) adding timestamps when the data is received by a debug adapter hardware between the SoC and the host PC.
Security and Authentication¶
Any debug system, by nature, exposes much of the system internals to the outside world. To prevent abuse of the debug system, production devices often require a developer to authenticate towards the system before being able to use the debug system.
OSD provides the infrastructure to implement such features.
OSD By Example¶
Before we dive into the details of the OSD architecture, this section discusses two typical usage examples of OSD. The first example only shows run-control debugging, the second one presents a full tracing infrastructure.
OSD for Run-Control Debugging¶
Many smaller single-core designs traditionally only support run-control debugging through custom JTAG-based debug infrastructure. OSD supports this use case well. Its modular architecture makes it easy to implement only essential debug modules to support run-control debug, and to add advanced features such as trace later without major changes.
Figure 1 shows an example configuration of OSD for a small run-control debug scenario. The functional system (to be attached on the right side) consists of a single-core CPU, a memory and a bus interconnect. To this functional system the debug modules are attached.
- The Subnet Control Module (SCM) module allows to control the system remotely: reset the system, halt the system, reset the CPUs, etc.
- The Core Debug Module (CDM) provides all functionality expected from a run-control debug system: setting breakpoints and reading CPU registers.
- The Memory Access Module (MAM) gives access to the chip’s memories for loading the memories during debugging (e.g. with the program code), to verify the memory contents, or to read out memory contents during debugging.
- To show the benefits of using OSD, the example system adds another module, the Device Emulation Module UART (DEM-UART). This module behaves on the functional hardware side, and on the software side like a usual UART device. But instead of using dedicated pins, the data is transported through the debug connection.
For all mentioned components, OSD includes a full specifications which enables a custom implementation, as well as a hardware implementation that can be used unmodified or adapted to fit the interface to the custom functional system.
The debug modules are all connected to a debug network. The OSD specification does not require a specific network topology or implementation type. However, usually OSD implementations use a 16-bit wide, unidirectional ring network on chip (NoC), as it presents a good trade-off between area usage and performance.
To connect with a host PC, three further components are needed: the Host Interface Module (HIM) on the hardware side, a GLIP transport module, and a software daemon on the host side.
The transport of data between host and device is handled by GLIP. GLIP is a library which abstracts the data transport between hardware and software with a bi-directional FIFO interface. The data transport itself can happen through different physical interfaces, such as UART, JTAG, USB or PCI Express (PCIe). In the presented example, a JTAG connection is used. A possibly existing JTAG boundary scan interface can be re-used and a new Test Access Point (TAP) is added to the JTAG chain for the debug connection.
The Host Interface Module (HIM) connects the debug network to the FIFO-interface of GLIP.
On the software side, the OSD host daemon encapsulates the communication to the device and provides a API for various tools communicating with the debug system. A scriptable command line interface can be used to control the system (such as reset, halt, etc.) and to read and write memories. A gdb server provides an interface to the core debug functionality that the GNU Debugger (gdb) can connect to. In the end, software can be debugged with an unmodified gdb (and other gdb-enabled IDEs, such as Eclipse CDT).
OSD for Tracing¶
Today’s debug system architectures strictly separate between run-control debugging and tracing. The example below shows how OSD units the two worlds with a common interface, thus reducing development and maintainance effort. Since most of the architecture is shared between run-control debugging and tracing, upgrading an existing design from run-control debugging to tracing is not a large step.
Figure 2 shows an example architecture of a OSD system with tracing support for a dual-core design. Most of the architecture is identical to the previous example: the host daemon, the HIM, the debug network and the SCM, CDM and MAM debug modules. New in this example are the following parts.
- The GLIP transport library now uses USB 2.0 instead of JTAG for communication. This allows for higher off-chip transfer speeds to get improved visibility into the system by tracing.
- The Core Trace Module (CTM) provides program trace (a.k.a. instruction trace) support. It is attached to the CPU core next to the CDM.
- A graphical trace viewer can be attached to the host daemon to view the traces. Currently, OSD does not come with such a tool, but all interfaces are provided to easily write such a tool.
The two examples in this section have already shed a light on what is possible with OSD. In the remainder of this document, we’ll discuss OSD in more depth, starting with a more general overview of the architecture.