Every hardware peripheral needs some form of higher-level software to actuate it to perform its intended function. Device Interface Functions (DIFs) aim to make it easy to use hardware for its intended purposes. DIFs can be seen as a working code reference for interacting with a given piece of hardware.
DIFs provide extensively reviewed APIs for actuating hardware for three separate use cases: design verification, FPGA + post-silicon validation, and providing example code to aid the implementation of non-production firmware. DIFs may be illustrative for writing device drivers but should not be considered drivers themselves.
DIFs must be written in C, specifically C11 (with a few allowed extensions). DIFs must conform to the style guide in sw/device/lib/dif.
DIFs must only depend on:
sw/device/lib/base/bitfield.h),sw/device/lib/base/mmio.h),sw/device/lib/base/memory.h), andDIFs must not depend on DIFs for other IP blocks, or other external libraries.
This decision is motivated by the requirement that DIFs must be extremely flexible in their possible execution environments, including being hosted in bare metal, or being called from other languages such as Rust through Foreign Function Interface.
DIFs must not depend on runtime services provided by an operating system. If functions must be called in response to system stimulus such as interrupts then the DIF must make these requirements explicit in their documentation. This decision makes DIFs appropriate for execution environments without OS support such as DV.
DIFs should not depend on architecture-specific constructs such as inline assembly or platform-defined registers, except where a peripheral is integrated into the core in a way making them unavoidable. As a concrete example: a SPI DIF must not depend on a pinmux or clock control DIF in order to be operated. This highlights a clear separation of concern: making a pin on the package driven from the SPI block involves multiple systems. By design, a DIF cannot coordinate cross-functional control.
DIFs must actuate all of the specification-required functionality of the hardware that they are written for, and no more. A DIF cannot be declared complete until it provides an API for accessing all of the functionality that is expected of the hardware. This is distinct from mandating that a DIF be required to cover all of the functionality of a given piece of hardware.
DV, FPGA, and early silicon validation have unique requirements in their use of hardware. They might actuate hardware on vastly different timescales, use verification-specific registers, or otherwise manipulate aspects of the hardware that “production”-level code would not (or cannot) do. To this end, verification-only functionality may be added to a DIF only in modules that are included in verification. This functionality must not be made accessible outside of verification environments, and is enforced by not including DV-specific code in non-DV builds.
In addition to the interface functions themselves, DIFs are expected to usually take the form of a structure that contains instance-specific information such as base register addresses and non-hardware-backed state required to actuate the hardware. DIFs should not track hardware-backed state in their own state machine.
A DIF instance is expressly an instance of a hardware block, so DIFs must not intrinsically know the location of the hardware blocks they're associated with in memory. That is:
dif_timer_init_result_t dif_timer_init(dif_timer_t *timer, const timer_config_t *config, uintptr_t base_address);
allows the system initialization code to instantiate timers at specific locations, whereas:
// Don't do this: bool dif_timer_init(timer_t *timer, enum timer_num num);
suggests that the timer DIF knows about the number and placement of timers in the system.
DIFs must not store or provide information in their implementation or state that is outside of their area of concern. The number of timers in a system is the concern of the system, not the timer DIF.
All DIF functions must have clear direct objects and verbs, be written in the imperative mood, and follow the format of dif_<object>_<verb_phrase>. The object in the name must be common to the DIF it appears it, and unique among DIFs.
Consider the following examples of good names:
dif_timer_init,dif_timer_reset,dif_timer_set_time_remaining,dif_timer_get_time_remaining.The following are bad names:
dif_clear_timer (wrong object/verb-phrase ordering),dif_timer_gets_reset (passive voice not imperative),timer_init (not correctly prefixed for C).Prefer common names: init is more commonly written than initialize, but reset is more commonly written than rst, for example. This is a subjective call on the DIF author and reviewers' parts, so common agreement will be more useful than strict prescription.
All DIF exported types and functions must have associated API documentation. All function parameters must be documented. All function semantics must be documented, no matter how obvious one suspects the function to be. The documentation should be exhaustive enough that the implementation can reasonably be inferred from it and the IP specification.
It is important that the DIFs are documented alongside the hardware IP that they are written for. Each hardware IP block also contains a “programmers' guide” section, which tells programmers how to use the hardware interfaces correctly, using prose descriptions and relevant diagrams. The prose descriptions should include references to relevant DIF functions.
Programmers' guides should primarily refer to the Software API Documentation. Programmers' guides should not contain standalone examples that do not match how we implement DIFs and can become out of date. The software API documentation includes full source code of the respective DIFs, should extra clarity be needed to explain how a hardware interface should be used.
The programmers' guide must also include a list of DIF functions which relate to it, along with a brief description of each, and a link to further API documentation.
DIFs must live in the sw/device/lib/dif folder. They must comprise at least one header unique to the hardware in question, and may contain more than one C file providing the implementation, if the complexity of the implementation warrants it (this is subjective). Any files for a given IP should have a filename starting dif_<IP name>.
Verification-specific extensions and logic must not live in sw/device/lib/dif. This code must be placed in the directory belonging to the verification domain which the extension is applicable to. Verification-specific extensions must not be built and included in production firmware builds.
Tests must live in the sw/device/tests/dif folder.
DIFs must have unit tests, which cover all their specification-required functionality. These unit tests use a mocking system which allows the test to control all hardware reads and writes, without requiring complex management of the hardware. These unit tests are written using googletest, and may be run on a non-RISC-V host.
DIFs should also have other tests, including standalone C sanity tests---which can quickly diagnose major issues between the DIF and the hardware but are not required to test all device functionality.
DIFs are also being used by DV for chip-level tests, which should help catch any issues with the DIFs not corresponding to the hardware implementation exactly. These tests may not live in sw/device/tests/dif.
As part of the DIF Standardisation process, we've decided to establish more concrete DIF lifecycle stages. These are documented with the Development Stages.
In the hardware world, there are checklists for establishing that each stage is complete and the next stage can be moved to.
In software, we don't have to follow the same stability guarantees, because software is much more modifiable. It makes sense to have some kind of review process, but this should be much more lightweight in the early stages and not significantly burdensome to the associated HW designer.
The current proposal only covers a DIF being written against a single version of its respective hardware IP block. This specifically excludes how to disambiguate DIFs in the repository that are written against different versions of the same IP block, or writing a single DIF that is compatible with multiple versions of an IP block.
The DIF lead author is expected to participate in the hardware IP block's L2 signoff review. This review can only happen once hardware design and verification are complete, and the DIF itself has reached stage S3.