The HAL Port

Name

HAL Port -- Implementation Details

Description

This documentation explains how the eCos HAL specification has been mapped onto M68K hardware, and should be read in conjunction with that specification. It also describes how variant, processor and platform HALs can modify the default behaviour.

eCos support for any given target will involve either three or four HAL packages: the architectural HAL, the platform HAL, the variant HAL, and optionally a processor HAL. This package, the architectural HAL, provides code and definitions that are applicable to all M68K processors. The platform HAL provides support for one specific board, for example an M5272C3 evaluation board, or possibly for a number of almost-identical boards. The processor HAL, if present, serves mainly to provide details of on-chip peripherals including the interrupt controller. The variant HAL provides functionality that is common to a group of processors, for example all MCFxxxx processors have very similar UARTs and hence can share HAL diagnostic code. There is no fixed specification of what should go into the variant HAL versus the processor HAL. For simplicity the description below only refers to variant HALs, but the work may actually happen in a processor HAL instead.

As a design goal lower-level HALs can always override functionality that is normally provided higher up. For example the architectural HAL will provide the required eCos HAL_LSBIT_INDEX and HAL_MSBIT_INDEX macros, but these can be provided lower down instead. Many but not all ColdFire processors have the ff1 and bitrev instructions which allow for a more efficient implementation than the default architectural ones. In some areas such as handling context switching the architectural HAL will usually provide the basic functionality but it may be extended by lower HALs, for example to add support for the multiply-accumulate units present in certain ColdFire processors.

The architectural HAL provides header files cyg/hal/hal_arch.h, cyg/hal/hal_intr.h, cyg/hal/hal_cache.h, cyg/hal/hal_io.h and cyg/hal/arch.inc. These automatically include an equivalent header file from the variant HAL, for example cyg/hal/var_arch.h. The variant HAL header will in turn include processor and platform-specific headers. This means that application developers and other packages can simply include the architectural HAL headers without needing to know about variants or platforms. It also allows the variant and platform HALs to override architectural settings.

The port assumes that eCos and application code always runs in supervisor mode, with full access to all hardware and special registers.

Data Types

For eCos purposes all M68K processors are big-endian and 32-bit, so the default data types in cyg/infra/cyg_type.h are used. Some variants have external bus widths less than 32-bit, but this does not affect the architectural HAL.

When porting to another variant it is possible to override some or all of the type definitions. The variant HAL needs to implement the CDL interface CYGINT_HAL_M68K_VARIANT_TYPES and provide a header file cyg/hal/var_basetype.h.

Startup and Exception Vectors

The conventional bootstrap mechanism involves a table of exception vectors at the base of memory. The first two words of this table give the initial program counter and stack pointer. In a typical embedded system the hardware is arranged such that non-volatile flash memory is found at location 0x0 so it is the start of flash that contains the exception vectors and the boot code. The table of exception vectors is used subsequently for interrupt handling and for hardware exceptions such as attempts to execute an illegal instruction. There are a number of common scenarios:

  1. On systems with very limited memory flash may remain mapped at location 0 and the table of exception vectors remains mapped there as well. The M68K architecture defines the table to have 256 entries and hence it occupies 1K of memory, but in reality many of the entries are unused so part of the table may get used for code instead. Since the whole exception vector table is in read-only memory parts of the eCos interrupt and exception handling mechanisms have to be statically initialized and macros like HAL_VSR_SET are not available.

  2. As a minor variation of the previous case, flash remains at location 0 but the table of exception vectors gets remapped elsewhere in the address space, usually RAM. This allows HAL_VSR_SET to operate normally but at the cost of increased memory usage. The exception vector table in flash only contains two entries, for the initial program counter and stack pointer. The exception vector table in RAM typically gets initialized at run-time.

  3. On systems with more memory it is conventional to rearrange the address map during bootstrap. The flash gets relocated, typically to near the end of the address space, and RAM gets placed at location 0 instead. The exception vector table stays at location 0 but is now in RAM and gets initialized at run-time. The bootstrap exception vector table in flash again only needs two entries. A variation places the RAM elsewhere in the address space and moves the exception vector table there, leaving location 0 unused. This provides some protection against null pointer accesses in errant code.

    As a further complication, larger systems typically support different startup types. The application can be linked against a ROM startup configuration and placed directly in flash, as before. Alternatively there may be a ROM monitor such as RedBoot in the flash, taking care of initial bootstrap. The user's application is linked against a RAM startup configuration, loaded into RAM via the ROM monitor, and debugged using functionality provided by the ROM monitor. Yet another possibility involves a RAM startup application but it gets loaded and run via a hardware debug technology such as BDM, and the ROM monitor is either missing or not used.

The exact hardware details, the various startup types, the steps needed for low-level hardware initialization, and so on are not known to the architectural HAL. Hence although the architectural HAL does provide the basic framework for startup, much of the work is done via macros provided by lower-level HAL packages and those macros are likely to depend on various configuration options. Rather than try to enumerate all the various combinations here it is better to look at the actual code in vectors.S and in appropriate variant, processor or platform HALs. vectors.S is responsible for any low-level initialization that needs to happen. This includes setting up a standard C environment with the stack pointer set to the startup stack in working RAM, making sure all statically initialized global variables have the correct values, and that all uninitialized global variables are zeroed. Once the C environment has been set up the code jumps to hal_m68k_c_startup in file hal_m68k.c which completes the initialization and jumps to the application entry point.

Interrupt Handling

The M68K architecture reserves a 1K area of memory for 256 exception vectors. These are used for internal and external interrupts, exceptions, software traps, and special operations such as reset handling. Some of the vectors have well-defined uses. However when it comes to interrupt handling the details will depend on the processor variant and on the platform, and the appropriate package documentation should be consulted for full details. Most platforms will not use the full set of 256 vectors, instead re-using some of this memory for other purposes.

By default the exception vectors are located at location 0, but some variants allow the vectors to be located elsewhere. This is managed by an M68K-specific macro CYG_HAL_VSR_TABLE. The default value is 0, but a variant HAL can provide an alternative value.

The standard eCos macros HAL_VSR_GET and HAL_VSR_SET just manipulate one of the 256 entries in the table of exception vectors. Hence it is usually possible to replace the default handlers for exceptions and traps in addition to interrupt handlers. hal_intr.h provides #define's for the more common exception vectors, and additional ones can be provided by the platform or variant. It is the responsibility of the platform or variant HAL to initialize the table, and to provide the HAL_VSR_SET_TO_ECOS_HANDLER macro since that requires knowledge of the default table entries.

It should be noted that in some configurations the table of exception vectors may reside in read-only memory so entries cannot be changed. If so then the HAL_VSR_SET and HAL_VSR_SET_TO_ECOS_HANDLER macros will not be defined. Portable code may need to consider this possibility and test for the existence of these macros before using them.

The architectural HAL provides an entry point hal_m68k_interrupt_vsr in the file hal_arch.S. When an interrupt occurs the original 68000 pushed the program counter and the status register on to the stack, and then called the VSR via the exception table. On newer variants some additional information is pushed, including details of the interrupt source. hal_m68k_interrupt_vsr assumes the latter and can be used directly as the VSR on these newer variants. On older variants a small trampoline is needed which pushes the additional information and then jumps to the generic VSR. Interpreting the additional information is handled via an assembler macro hal_context_extract_isr_vector_shl2 which should be defined by the variant, matching the behaviour of the hardware or the trampoline.

At the architecture level there is no fixed mapping between VSR and ISR vectors. Instead that is left to the variant or platform HAL. The architectural HAL does provide default implementations of HAL_INTERRUPT_ATTACH, HAL_INTERRUPT_DETACH and HAL_INTERRUPT_IN_USE since these just involve updating a static table.

By default the interrupt state control macros HAL_DISABLE_INTERRUPTS, HAL_RESTORE_INTERRUPTS, HAL_ENABLE_INTERRUPTS and HAL_QUERY_INTERRUPTS are implemented by the architectural HAL, and simply involve updating the status register. Disabling interrupts involves setting the three IPL bits to 0x07. Enabling interrupts involves setting those bits to a smaller value, CYGNUM_HAL_INTERRUPT_DEFAULT_IPL_LEVEL, which defaults to 0.

HAL_DISABLE_INTERRUPTS has no effect on non-maskable interrupts. This causes a problem because parts of the system assume that all normal interrupt sources are affected by this macro. If the target hardware can raise non-maskable interrupts then it is the responsibility of application code to install a suitable VSR and handle non-maskable interrupts entirely within the application, bypassing the usual eCos ISR and DSR mechanisms.

The architectural HAL does not provide any support for the interrupt controller management macros like HAL_INTERRUPT_MASK. These can only be implemented on a per-variant, per-processor or per-platform basis.

Exception Handling

Synchronous exception handling is done in much the same way as interrupt handling. The architectural HAL provides a generic entry point hal_m68k_exception_vsr. On some variants this can be used directly as the exception VSR, on others it will be called via a small trampoline.

The details of exception handling vary widely from one variant to the next. Some variants push a great deal of additional information on to the stack for certain exceptions, but not all. The pushed program counter may correspond to the specific instruction that caused the exception, or the next instruction, or there may be only a loose correlation because of buffered writes. The architectural HAL makes no attempt to cope with all these differences, although some variants may provide more advanced support. Otherwise if an exception needs to be handled in a very specific way then it is up to the application to install a suitable VSR and handle the exception directly.

Stacks and Stack Sizes

cyg/hal/hal_arch.h defines values for minimal and recommended thread stack sizes, CYGNUM_HAL_STACK_SIZE_MINIMUM and CYGNUM_HAL_STACK_SIZE_TYPICAL. These values are specific to the current configuration, and are affected mainly by options related to interrupt handling.

By default eCos uses a separate interrupt stack, although this can be disabled through the configuration option CYGIMP_HAL_COMMON_INTERRUPTS_USE_INTERRUPT_STACK. When an interrupt or exception occurs eCos will save the context on the current stack and then switch to the interrupt stack before calling the appropriate ISR interrupt handler. This means that thread stacks can be significantly smaller because there is no need to worry about interrupt handling overheads, just the thread context. However switching the stack does require some extra work and hence increases the interrupt latency. Disabling the interrupt stack removes this processing overhead but requires larger stack sizes. It depends on the application whether or not this is a sensible trade off.

By default eCos does not allow nested interrupts, but this can be controlled via the configuration option CYGSEM_HAL_COMMON_INTERRUPTS_ALLOW_NESTING. Supporting nested interrupts requires larger thread stacks, especially if the separate interrupt stack is also disabled.

Although the M68K has enough registers for typical operation, the calling conventions are memory-oriented. In particular all arguments are pushed on the stack rather than held in registers, and the return address is also pushed rather than ending up in a link register. To allow for this the recommended minimum stack sizes are a little bit larger than for some other architectures. Variant HALs cannot directly affect these stack sizes. However the sizes do depend in part on the size of a thread context, so if for example the processor supports hardware floating point and support for that is enabled then the stack sizes will increase.

Usually the M68K architectural HAL will provide a single block of memory which acts as both the startup and interrupt stack, and there are configuration options to control the size of this block. Alternatively a variant, processor or platform HAL may define either or both of _HAL_M68K_STARTUP_STACK_ and _HAL_M68K_INTERRUPT_STACK_BASE_ if for some reason the stacks should not be placed in ordinary RAM.

Thread Contexts and Setjmp/Longjmp

A typical thread context consists of the following:

  1. The integer context. This consists of the data registers %d0 to %d7 and the address registers %a0 to %a6, The stack pointer register %a7 does not have to be saved explicitly since it is implicit in the pointer to the saved context.

    The caller-save registers are %d0, %d1, %a0, %a1, %a7 and the status register. The remaining registers are callee-save. Function arguments are always passed on the stack. The result is held in %d0.

  2. Floating point context, consisting of eight 64-bit floating point registers %fp0 to %fp7 and two support registers %fpsr and %fpiar. Support for this is only relevant if the processor variant has a hardware floating point unit, and even then saving floating point context is optional and can be disabled using a configuration option CYGIMP_HAL_M68K_FPU_SAVE. The control register %fpcr is not saved as part of the context. It is assumed that a single %fpcr value, usually 0, will be used throughout the application.

    The architectural HAL provides support for the hardware floating point unit. The variant or processor HAL should implement the CDL interface CYGINT_HAL_M68K_VARIANT_FPU if this hardware unit is actually present.

  3. Some M68K variants have additional hardware units, for example the multiply-accumulate units in certain ColdFire processors. The architectural HAL allows the context to be extended through various macros such as HAL_CONTEXT_OTHER.

  4. The status register %sr and the program counter. These are special because when an interrupt occurs the hardware automatically pushes these onto the stack, but exactly what gets pushed depends on the variant.

setjmp and longjmp only deal with the integer and fpu contexts. It is assumed that any special hardware units will only be used by application code, not by the compiler. Hence it is the responsibility of application code to define and implement appropriate setjmp semantics for these units. The variant HAL package can override the default implementations if necessary.

When porting to a new M68K variant, if this has a hardware floating point unit then the variant HAL should implement the CDL interface CYGINT_HAL_M68K_VARIANT_FPU, thus enabling support provided by the architectural HAL. If the variant has additional hardware units involving state that should be preserved during a context switch or when an interrupt occurs, the variant HAL should define a number of macros. The header file cyg/hal/var_arch.h should define HAL_CONTEXT_OTHER, HAL_CONTEXT_OTHER_SIZE, and HAL_CONTEXT_OTHER_INIT, either directly or via cyg/hal/proc_arch.h. The assembler header file cyg/hal/var.inc should define a number of macros such as hal_context_other_save_caller. For details of these macros see the architectural hal_arch.S file.

Variants also need to define exactly how the status register and program counter are saved onto the stack when an interrupt or exception occurs. This is handled through C macros HAL_CONTEXT_PCSR_SIZE, HAL_CONTEXT_PCSR_RTE_ADJUST, and HAL_CONTEXT_PCSR_INIT, and a number of assembler macros such as hal_context_pcsr_save_sr. Again the architectural files cyg/hal/hal_arch.h and hal_arch.S provide more details of these.

Bit Indexing

For performance reasons the HAL_LSBIT_INDEX and HAL_MSBIT_INDEX macros are implemented using assembler functions. A variant HAL can override the default definitions if, for example, the variant has special instructions to perform these operations.

Idle Thread Processing

The default HAL_IDLE_THREAD_ACTION implementation is a no-op. A variant HAL may override this, for example to put the processor into sleep mode. Alternative implementations should consider exactly how this macro gets used in eCos kernel code.

Clock Support

The architectural HAL cannot provide the required clock support because it does not know what timer hardware may be available on the target hardware. Instead this is left to either the variant or platform HAL, depending on whether the processor has a suitable on-chip timer or whether an off-chip timer has to be used.

HAL I/O

The M68K architecture does not have a separate I/O bus. Instead all hardware is assumed to be memory-mapped. Further it is assumed that all peripherals on the memory bus are wired appropriately for a big-endian processor and that there is no need for any byte swapping. Hence the various HAL macros for performing I/O simply involve pointers to volatile memory.

The variant, processor and platform equivalents of the cyg/hal/hal_io.h header will typically also provide details of some or all of the peripherals, for example register offsets and the meaning of various bits in those registers.

Cache Handling

If the processor has a cache then the variant HAL should implement the CDL interface CYGINT_HAL_M68K_VARIANT_CACHE. This causes the architectural header cyg/hal/hal_cache.h to pick up appropriate definitions from cyg/hal/var_cache.h. The architectural header will provide null defaults for anything not defined by the variant.

Linker Scripts

The architectural HAL will generate the linker script for eCos applications. This involves the architectural file m68k.ld and a .ldi memory layout file provided lower down, typically by the platform HAL. It is the LDI file which specifies the types and amount of memory available and which places code and data in appropriate places, but most of the hard work is done via macros provided by the architectural m68k.ld file.

Diagnostic Support

The architectural HAL does not implement diagnostic support. Instead this is left to the variant or platform HAL, depending on whether suitable peripherals are available on-chip or off-chip.

SMP Support

The M68K port does not have SMP support.

Debug Support

The M68K architectural HAL package provides basic support only for gdb stubs. There is no support for more advanced debug features like hardware watchpoints.

The generic gdb support in the common HAL requires a platform header <cyg/hal/plf_stub.h. In practice there is rarely any need for the contents of this file to change between platforms so the architectural HAL can provide a suitable default. It will do so if the CDL interface CYGINT_HAL_M68K_USE_STANDARD_PLATFORM_STUB_SUPPORT is implemented.

HAL_DELAY_US Macro

The architectural HAL provides a default implementation of the standard HAL_DELAY_US macro using a simply busy loop. To use this support a lower-level HAL should define _HAL_M68K_DELAY_US_LOOPS_, typically a small number of about 20 but it will need to be calibrated during the porting process. If the processor has a cache then the lower-level HAL may also define _HAL_M68K_DELAY_US_LOOPS_UNCACHED_ for the case when a delay loop is triggered while the cache is disabled.

Profiling Support

The M68K architectural HAL implements the mcount function, allowing profiling tools like gprof to determine the application's call graph. It does not implement the profiling timer. Instead that functionality needs to be provided by the variant or platform HAL. The implementation of mcount requires a dedicated frame pointer register so code should be compiled without the -fomit-frame-pointer flag.

Other Functionality

The M68K architectural HAL only implements the functionality provided by the eCos HAL specification and does not export any extra functionality.