Re: [flashv2 merge] io/flash

[Bart Veer <bartv at ecoscentric dot com>]

Switching to ecos-devel:

>>>>> "Jifl" == Jonathan Larmour <jifl@eCosCentric.com> writes:

    <snip>
    
    Jifl> To which Bart validly replied:
    Jifl> -=-=-=-=-=-=-
    Jifl> I raised it with Andrew back in July. Unfortunately it is
    Jifl> not currently possible. The currently defined constructor
    Jifl> priorities are not sufficiently flexible to cope with
    Jifl> dataflash, where the flash subsystem cannot be initialized
    Jifl> until after the SPI bus. It would be necessary to reorganize
    Jifl> the defined priorities, which is not something to be tackled
    Jifl> lightly.
    Jifl> -=-=-=-=-=-=-

    Jifl> Unfortunately discussion fizzled out. My bad.

I am not sure this is the right time to change the init priorities,
but on the other hand there is never going to be a perfect time for
that so we can at least have the discussion.

Let's start with the current priorities:

    #define CYG_INIT_HAL                    10000
    #define CYG_INIT_SCHEDULER              11000
    #define CYG_INIT_INTERRUPTS             12000
    #define CYG_INIT_DRIVERS                13000
    #define CYG_INIT_CLOCK                  14000
    #define CYG_INIT_IDLE_THREAD            15000
    #define CYG_INIT_THREADS                16000
    #define CYG_INIT_KERNEL                 40000
    #define CYG_INIT_MEMALLOC               47000
    #define CYG_INIT_IO                     49000
    #define CYG_INIT_IO_FS                  50000
    #define CYG_INIT_LIBC                   52000
    #define CYG_INIT_COMPAT                 55000
    #define CYG_INIT_APPLICATION            60000
    #define CYG_INIT_PREDEFAULT             65534
    #define CYG_INIT_DEFAULT                65535

I am not too worried about anything earlier than MEMALLOC. It
certainly makes sense to have the kernel fully initialized before
anything I/O related so that device drivers can register interrupt
handlers, create threads and alarms, malloc() memory, etc. I am also
not worried about anything that happens from LIBC onwards. However
INIT_IO and INIT_IO_FS are way too restrictive.

Now, I suspect that no solution is going to work for every bit of
hardware that people can conceive off. As and when we run into
problems we may have to add configury for setting certain init
priorities, together with requires constraints in the platform HAL, to
ensure that everything initializes in the correct order for a specific
platform. However the following order should work for most systems:

  #define CYG_INIT_BUS_PRIMARY
  #define CYG_INIT_BUS_PCI		(alias for PRIMARY)
  #define CYG_INIT_BUS_SECONDARY
  #define CYG_INIT_BUS_USBHOST		(alias for secondary)
  #define CYG_INIT_BUS_TERTIARY
  #define CYG_INIT_BUS_SPI		(alias for tertiary)
  #define CYG_INIT_BUS_I2C		(ditto)
  #define CYG_INIT_BUS_CAN		(?)
  #define CYG_INIT_DEV_WATCHDOG
  #define CYG_INIT_DEV_WALLCLOCK
  #define CYG_INIT_DEV_BLOCK_PRIMARY
  #define CYG_INIT_DEV_FLASH		(alias for BLOCK_PRIMARY)
  #define CYG_INIT_CONFIG
  #define CYG_INIT_DEV_BLOCK_SECONDARY
  #define CYG_INIT_DEV_CHAR		(all other devices)
  #define CYG_INIT_IO_FS

So we start by initializing the buses, in a hierarchical order.
Something like PCI has to be done first because subsequent drivers may
need to search for devices. Then comes USB host or similar, possibly
on a USB bus, because there may be slave devices hanging off the USB
host (although in practice I am not sure how usable a USB host
subsystem will be until the scheduler starts running). Finally we come
to the simpler buses like I2C and SPI. I am not sure whether CAN fits
into this category  or whether we should treat it more as a
network/char device.

Watchdog should probably be next so that if a subsequent device driver
init routine hangs the system can recover. Arguably watchdog should
happen earlier, right after kernel init, but that would cause problems
if you have an external watchdog on a bus rather than an on-chip one.
Right now the discussion is moot because watchdog devices are not
supposed to start automatically, instead an application call is
needed.

Doing the wallclock next is harmless and ensures that subsequent
operations have an accurate time stamp.

Now we come to primary block devices, primarily flash chips. SPI and
I2C have been initialized so there is no problem with serial flashes.

>From this point on we should have a non-volatile block device capable
of holding per-unit settings, so we can imagine an fconfig
implementation or replacement that initializes now and provides config
settings like MAC addresses to subsequent device drivers.

Another round of block devices, this time for things like disks.
Configuration data is available if needed. I am not entirely sure
about the need for this or about the order within the init sequence.

Then character devices, i.e. everything else including ethernet. I am
not sure there is any point in providing finer-grained control for
these.

Finally the file I/O subsystem. Possibly this should happen earlier,
between DEV_BLOCK_PRIMARY and CONFIG, so that an implementation of the
config data module can be layered on top of file I/O. Or possibly
CYG_INIT_IO_FS should happen immediately after CYG_INIT_MEMALLOC, with
the proviso that file I/O operations for devices may fail until later
in the init sequence.


Back to the original email and what should happen in the flash code,
off the top of my head my preference would be along the following
lines:

1) assume that diag_printf() is available after CYG_INIT_HAL, although
on a few platforms it may discard data until later in the init
sequence.

2) change the CHATTER() macro to cope with a NULL pf field.

3) default all devices to pf=NULL, which should be happening anyway
because the CYG_FLASH_DRIVER() macro does not initialize it.

4) remove the pf argument from cyg_flash_init() completely, so that by
default the flash subsystem is silent.

5) add a cyg_flash_set_printf() function which installs a
printf()-style function for a given flash device. Or possibly two
functions, one for a given device, another for all devices. Allow this
function to be called before cyg_flash_init().

6) change RedBoot to install diag_printf() before cyg_flash_init().

I think this automatically gives us the desired default behaviour for
both applications and RedBoot, and may end up removing a diag_printf()
dependency for some applications.

Or we could be more ambitious and get rid of cyg_flash_init()
completely, replacing it with a CYG_INIT_DEV_FLASH constructor.

Bart

-- 
Bart Veer                                   eCos Configuration Architect
eCosCentric Limited    The eCos experts      http://www.ecoscentric.com/
Barnwell House, Barnwell Drive, Cambridge, UK.      Tel: +44 1223 245571
Registered in England and Wales: Reg No 4422071.