Instantiating — including the driver in an eCos target


#include <cyg/io/am29xxxxx_dev.h>

int cyg_am29xxxxx_init_check_devid_XX(struct cyg_flash_dev* device);

int cyg_am29xxxxx_init_cfi_XX(struct cyg_flash_dev* device);

int cyg_am29xxxxx_erase_XX(struct cyg_flash_dev* device, cyg_flashaddr_t addr);

int cyg_am29xxxxx_program_XX(struct cyg_flash_dev* device, cyg_flashaddr_t addr, const void* data, size_t len);

int cyg_at49xxxx_softlock(struct cyg_flash_dev* device, const cyg_flashaddr_t addr);

int cyg_at49xxxx_hardlock(struct cyg_flash_dev* device, const cyg_flashaddr_t addr);

int cyg_at49xxxx_unlock(struct cyg_flash_dev* device, const cyg_flashaddr_t addr);

int cyg_am29xxxxx_read_devid_XX(struct cyg_flash_dev* device);


The AM29xxxxx family contains some hundreds of different flash devices, all supporting the same basic set of operations but with various common or uncommon extensions. The devices vary in capacity, performance, boot block layout, and width. There are also platform-specific issues such as how many devices are actually present on the board and where they are mapped in the address space. The AM29xxxxx driver package cannot know the details of every chip and every platform. Instead it is the responsibility of another package, usually the platform HAL, to supply the necessary information by instantiating some data structures. Two pieces of information are especially important: the bus configuration and the boot block layout.

Flash devices are typically 8-bits, 16-bits, or 32-bits wide (64-bit devices are not yet in common use). Most 16-bit devices will also support 8-bit accesses, but not all. Similarly 32-bit devices can be accessed 16-bits at a time or 8-bits at a time. A board will have one or more of these devices on the bus. For example there may be a single 16-bit device on a 16-bit bus, or two 16-bit devices on a 32-bit bus. The processor's bus logic determines which combinations are possible, and there will be a trade off between cost and performance: two 16-bit devices in parallel can provide twice the memory bandwidth of a single device. The driver supports the following combinations:

A single 8-bit flash device on an 8-bit bus.
A single 16-bit flash device on a 16-bit bus.
A single 32-bit flash device on an 32-bit bus.
Two parallel 8-bit devices on an 16-bit bus.
Four parallel 8-bit devices on a 32-bit bus.
Two parallel 16-bit devices on a 32-bit bus, with one device providing the bottom two bytes of each 32-bit datum and the other device providing the top two bytes.
A single 16-bit flash device connected to an 8-bit bus.
A single 32-bit flash device connected to a 16-bit bus.

These configuration all require slightly different code to manipulate the hardware. The AM29xxxxx driver package provides separate functions for each configuration, for example cyg_am29xxxxx_erase_16 and cyg_am29xxxxx_program_1616.


At the time of writing not all the configurations have been tested.

The second piece of information is the boot block layout. Flash devices are subdivided into blocks (also known as sectors - both terms are in common use). Some operations such as erase work on a whole block at a time, and for most applications a block is the smallest unit that gets updated. A typical block size is 64K. It is inefficient to use an entire 64K block for small bits of configuration data and similar information, so many flash devices also support a number of smaller boot blocks. A typical 2MB flash device could have a single 16K block, followed by two 8K blocks, then a 32K block, and finally 31 full-size 64K blocks. The boot blocks may appear at the bottom or the top of the device. So-called uniform devices do not have boot blocks, just full-size ones. The driver needs to know the boot block layout. With modern devices it can work this out at run-time, but often it is better to provide the information statically.


In most cases flash support is specific to a platform. Even if two platforms happen to use the same flash device there are likely to be differences such as the location in the address map. Hence there is little possibility of re-using the platform-specific code, and this code should be placed in the platform HAL rather than in a separate package. Typically this involves a separate file and a corresponding compile property in the platform HAL's CDL:

cdl_package CYGPKG_HAL_M68K_ALAIA {
    compile -library=libextras.a alaia_flash.c

The contents of this file will not be accessed directly, only indirectly via the generic flash API, so normally it would be removed by link-time garbage collection. To avoid this the object file has to go into libextras.a.

The actual file alaia_flash.c will look something like:

#include <pkgconf/system.h>

#include <cyg/io/flash.h>
#include <cyg/io/flash_dev.h>
#include <cyg/io/am29xxxxx_dev.h>

static const CYG_FLASH_FUNS(hal_alaia_flash_amd_funs,
    (int (*)(struct cyg_flash_dev*, const cyg_flashaddr_t, void*, size_t))0,

static const cyg_am29xxxxx_dev hal_alaia_flash_priv = {
    .devid      = 0x45,
    .block_info = {
        { 0x00004000, 1 },
        { 0x00002000, 2 },
        { 0x00008000, 1 },
        { 0x00010000, 63 }


The bulk of the file is protected by an #ifdef for the AM29xxxxx flash driver. That driver will only be active if the generic flash support is enabled. Without that support there will be no way of accessing the device so instantiating the data structures would serve no purpose. The rest of the file is split into three structure definitions. The first supplies the functions which will be used to perform the actual flash accesses, using a macro provided by the generic flash code in cyg/io/flash_dev.h. The relevant ones have an _16 suffix, indicating that on this board there is a single 16-bit flash device on a 16-bit bus. The second provides information specific to AM29xxxxx flash devices. The third provides the cyg_flash_dev structure needed by the generic flash code, which contains pointers to the previous two.


All eCos flash device drivers must implement a standard interface, defined by the generic flash code CYGPKG_IO_FLASH. This interface includes a table of seven function pointers for various operations: initialization, query, erase, program, read, locking and unlocking. The query operation is optional and the generic flash support provides a dummy implementation cyg_flash_devfn_query_nop. AM29xxxxx flash devices are always directly accessible so there is no need for a separate read function. The remaining functions are more complicated.

Usually the table can be declared const. In a ROM startup application this avoids both ROM and RAM copies of the table, saving a small amount of memory. const should not be used if the table may be modified by a platform-specific initialization routine.


There is a choice of three main initialization functions. The simplest is cyg_flash_devfn_init_nop, which does nothing. It can be used if the cyg_am29xxxxx_dev and cyg_flash_dev structures are fully initialized statically and the flash will just work without special effort. This is useful if it is guaranteed that the board will always be manufactured using the same flash chip, since the nop function involves the smallest code size and run-time overheads.

The next step up is cyg_am29xxxxx_init_check_devid_XX, where XX will be replaced by the suffix appropriate for the bus configuration. It is still necessary to provide all the device information statically, including the devid field in the cyg_am29xxxxx_dev structure. This initialization function will attempt to query the flash device and check that the provided device id matches the actual hardware. If there is a mismatch the device will be marked uninitialized and subsequent attempts to manipulate the flash will fail.

If the board may end up being manufactured with any of a number of different flash chips then the driver can perform run-time initialization, using a cyg_am29xxxxx_init_cfi_XX function. This queries the flash device as per the Common Flash Memory Interface Specification, supported by all current devices (although not necessarily by older devices). The block_info field in the cyg_am29xxxxx_dev structure and the end and num_block_infos fields in the cyg_flash_dev structure will be filled in. It is still necessary to supply the start field statically since otherwise the driver will not know how to access the flash device. The main disadvantage of using CFI is that it increases the code size.


If CFI is used then the cyg_am29xxxxx_dev structure must not be declared const. The CFI code will attempt to update the structure and will fail if the structure is held in read-only memory. This would leave the flash driver non-functional.

A final option is to use a platform-specific initialization function. This may be useful if the board may be manufactured with one of a small number of different flash devices and the platform HAL needs to adapt to this. The AM29xxxxx driver provides a utility function to read the device id, cyg_am29xxxxx_read_devid_XX:

static int
alaia_flash_init(struct cyg_flash_dev* dev)
    int devid = cyg_am29xxxxx_read_devid_1616(dev);
    switch(devid) {
        case 0x0042 :
        case 0x0084 :
          return CYG_FLASH_ERR_DRV_WRONG_PART;

There are many other possible uses for a platform-specific initialization function. For example initial prototype boards might have only supported 8-bit access to a 16-bit flash device rather than 16-bit access, but this problem was fixed in the next revision. The platform-specific initialization function can figure out which model board it is running on and replace the default 16as8 functions with faster 16 ones.

Erase and Program

The AM29xxxxx driver provides erase and program functions appropriate for the various bus configurations. On most targets these can be used directly. On some targets it may be necessary to do some extra work before and after the erase and program operations. For example if the hardware has an MMU then the part of the address map containing the flash may have been set to read-only, in an attempt to catch spurious memory accesses. Erasing or programming the flash requires write-access, so the MMU settings have to be changed temporarily. As another example some flash device may require a higher voltage to be applied during an erase or program operation. or a higher voltage may be desirable to make the operation proceed faster. A typical platform-specific erase function would look like this:

static int
alaia_flash_erase(struct cyg_flash_dev* dev, cyg_flashaddr_t addr)
    int result;
    …  // Set up the hardware for an erase
    result = cyg_am29xxxxx_erase_32(dev, addr);
    …  // Revert the hardware change
    return result;

There are two configurations which affect the erase and program functions, and which a platform HAL may wish to change: CYGNUM_DEVS_FLASH_AMD_AM29XXXXX_V2_ERASE_TIMEOUT and CYGNUM_DEVS_FLASH_AMD_AM29XXXXX_V2_PROGRAM_TIMEOUT. The erase and program operations both involve polling for completion, and these timeout impose an upper bound on the polling loop. Normally these operations should never take anywhere close to the timeout period, so a timeout indicates a catastrophic failure that should really be handled by a watchdog reset. A reset is particularly appropriate because there will be no clean way of aborting the flash operation. The main reason for the timeouts is to help with debugging when porting to new hardware. If there is a valid reason why a particular platform needs different timeouts then the platform HAL's CDL can require appropriate values for these options.


There is no single way of implementing the block lock and unlock operations on all AM29xxxxx devices. If these operations are supported at all then usually they involve manipulating the voltages on certain pins. This would not be able to be handled by generic driver code since it requires knowing how these pins can be manipulated via the processor's GPIO lines. Therefore the AM29xxxxx driver does not usually provide lock and unlock functions, and instead the generic dummy functions cyg_flash_devfn_lock_nop and cyg_flash_devfn_unlock_nop should be used. An exception exists for the AT49xxxx family of devices which are sufficiently AMD compatible in other respects. Otherwise, if a platform does provide a way of implementing the locking then this can be handled by platform-specific functions.

static int
alaia_lock(struct cyg_flash_dev* dev, const cyg_flashaddr_t addr)

static int
alaia_unlock(struct cyg_flash_dev* dev, const cyg_flashaddr_t addr)

If real locking functions are implemented then the platform HAL's CDL script should implement the CDL interface CYGHWR_IO_FLASH_BLOCK_LOCKING. Otherwise the generic flash package may believe that none of the flash drivers in the system provide locking functionality and disable the interface functions.

AT49xxxx locking

As locking is standardised across the AT49xxxx family of AMD AM29xxxxx compatible Flash parts, a method supporting this is included within this driver. cyg_at49xxxx_softlock_XX provides a means of locking a Flash sector such that it may be subsequently unlocked. cyg_at49xxxx_hardlock_XX locks a sector such that it cannot be unlocked until after reset or a power cycle. cyg_at49xxxx_unlock_XX unlocks a sector that has previously been softlocked. At power on or Flash device reset, all sectors default to being softlocked.


The driver provides a set of functions cyg_am29xxxxx_read_devid_XX, one per supported bus configuration. These functions take a single argument, a pointer to the cyg_flash_dev structure, and return the chip's device id. For older devices this id is a single byte. For more recent devices the id is a 3-byte value, 0x7E followed by a further two bytes that actually identify the device. cyg_am29xxxxx_read_devid_XX is usually called only from inside a platform-specific driver initialization routine, allowing the platform HAL to adapt to the actual device present on the board.

Device-Specific Structure

The cyg_am29xxxxx_dev structure provides information specific to AM29xxxxx flash devices, as opposed to the more generic flash information which goes into the cyg_flash_dev structure. There are only two fields: devid and block_info.

devid is only needed if the driver's initialization function is set to cyg_am29xxxxx_init_check_devid_XX. That function will extract the actual device info from the flash chip and compare it with the devid field. If there is a mismatch then subsequent operations on the device will fail.

The block_info field consists of one or more pairs of the block size in bytes and the number of blocks of that size. The order must match the actual hardware device since the flash code will use the table to determine the start and end locations of each block. The table can be initialized in one of three ways:

  1. If the driver initialization function is set to cyg_flash_devfn_init_nop or cyg_am29xxxxx_init_check_devid_XX then the block information should be provided statically. This is appropriate if the board will also be manufactured using the same flash chip.
  2. If cyg_am29xxxxx_init_cfi_XX is used then this will fill in the block info table. Hence there is no need for static initialization.
  3. If a platform-specific initialization function is used then either this should fill in the block info table, or the info should be provided statically.

The size of the block_info table is determined by the configuration option CYGNUM_DEVS_FLASH_AMD_AM29XXXXX_V2_ERASE_REGIONS. This has a default value of 4, which should suffice for nearly all AM29xxxxx flash devices. If more entries are needed then the platform HAL's CDL script should require a larger value.

If the cyg_am29xxxxx_dev structure is statically initialized then it can be const. This saves a small amount of memory in ROM startup applications. If the structure is updated at run-time, either by cyg_am29xxxxx_init_cfi_XX or by a platform-specific initialization routine, then it cannot be const.

Flash Structure

Internally the generic flash code works in terms of cyg_flash_dev structures, and the platform HAL should define one of these. The structure should be placed in the cyg_flashdev table. The following fields need to be provided:

This should point at the table of functions.
The base address of the flash in the address map. On some board the flash may be mapped into memory several times, for example it may appear in both cached and uncached parts of the address space. The start field should correspond to the cached address.
The address of the last byte in the flash. It can either be statically initialized, or cyg_am29xxxxx_init_cfi_XX will calculate its value at run-time.
This should be the number of entries in the block_info table. It can either be statically initialized or it will be filled in by cyg_am29xxxxx_init_cfi_XX.
The table with the block information is held in the cyg_am29xxxxx_dev structure, so this field should just point into that structure.
This field is reserved for use by the device driver. For the AM29xxxxx driver it should point at the appropriate cyg_am29xxxxx_dev structure.

The cyg_flash_dev structure contains a number of other fields which are manipulated only by the generic flash code. Some of these fields will be updated at run-time so the structure cannot be declared const.

Multiple Devices

A board may have several flash devices in parallel, for example two 16-bit devices on a 32-bit bus. It may also have several such banks to increase the total amount of flash. If each device provides 2MB, there could be one bank of 2 parallel flash devices at 0xFF800000 and another bank at 0xFFC00000, giving a total of 8MB. This setup can be described in several ways. One approach is to define two cyg_flash_dev structures. The table of function pointers can usually be shared, as can the cyg_am29xxxxx_dev structure. Another approach is to define a single cyg_flash_dev structure but with a larger block_info table, covering the blocks in both banks of devices. The second approach makes more efficient use of memory.

Many variations are possible, for example a small slow flash device may be used for initial bootstrap and holding the configuration data, while there is also a much larger and faster device to hold a file system. Such variations are usually best described by separate cyg_flash_dev structures.

If more than one cyg_flash_dev structure is instantiated then the platform HAL's CDL script should implement the CDL interface CYGHWR_IO_FLASH_DEVICE once for every device past the first. Otherwise the generic code may default to the case of a single flash device and optimize for that.

Platform-Specific Macros

The AM29xxxxx driver source code includes the header files cyg/hal/hal_arch.h and cyg/hal/hal_io.h, and hence indirectly the corresponding platform header files (if defined). Optionally these headers can define macros which are used inside the driver, thus giving the HAL limited control over how the driver works.

Cache Management

By default the AM29xxxxx driver assumes that the flash can be accessed uncached, and it will use the HAL CYGARC_UNCACHED_ADDRESS macro to map the cached address in the start field of the cyg_flash_dev structure into an uncached address. If for any reason this HAL macro is inappropriate for the flash then an alternative macro HAL_AM29XXXXX_UNCACHED_ADDRESS can be defined instead. However fixing the CYGARC_UNCACHED_ADDRESS macro is normally the better solution.

If there is no way of bypassing the cache then the platform HAL should implement the CDL interface CYGHWR_DEVS_FLASH_AMD_AM29XXXXX_V2_CACHED_ONLY. The flash driver will now disable and re-enable the cache as required. For example a program operation will involve the following:

while ( ! finished ) {

The default implementations of these INTSCACHE macros are as follows: STATE defines any local variables that may be needed, e.g. to save the current interrupt state; BEGIN disables interrupts, synchronizes the data caches, disables it, and invalidates the current contents; SUSPEND re-enables the data cache and then interrupts; RESUME disables interrupts and the data cache; END re-enables the cache and then interrupts. The cache is only disabled when interrupts are disabled, so there is no possibility of an interrupt handler running or a context switch occurring while the cache is disabled, potentially leaving the system running very slowly. The data cache synchronization ensures that there are no dirty cache lines, so when the cache is disabled the low-level flash write code will not see stale data in memory. The invalidate ensures that at the end of the operation higher-level code will not pick up stale cache contents instead of the newly written flash data. The SUSPEND and RESUME macros only re-enable and disable the data cache. An interrupt and possibly a context switch may occur between these macros and use the cache normally. It is assumed that any code which runs at this time will not touch the memory being used by the flash operation, so as far as the low-level program code is concerned it can just continue to use the uncached memory contents as set up by the BEGIN macro. If any code modifies the const data currently being written to a flash block or tries to read the flash block being modified then the system's behaviour is undefined. Theoretically a more robust approach is possible, synchronizing and invalidating the cache again in every RESUME. However these cache operations can be expensive and RESUME may get invoked some thousands of times for every flash block, so this alternative approach would cripple the driver's performance.

Some implementations of the HAL cache macros may not provide the exact semantics required by the flash driver. For example HAL_DCACHE_DISABLE may have an unwanted side effect, or it may do more work than is needed here. The driver will check for alternative macros HAL_AM29XXXXX_INTSCACHE_STATE, HAL_AM29XXXXX_INTSCACHE_BEGIN, HAL_AM29XXXXX_INTSCACHE_SUSPEND, HAL_AM29XXXXX_INTSCACHE_RESUME and HAL_AM29XXXXX_INTSCACHE_END, using these instead of the defaults.