[U-Boot] [PATCH] [WIP] sandbox: spi/sf emulation

[Simon Glass]

Hi Mike,

On Fri, Oct 28, 2011 at 5:56 PM, Mike Frysinger <vapier at gentoo.org> wrote:
> On Sat, Oct 29, 2011 at 02:47, Simon Glass wrote:
>> On Wed, Oct 26, 2011 at 1:37 PM, Mike Frysinger wrote:
>>> On Wed, Oct 26, 2011 at 12:26, Simon Glass wrote:
>>>> On Wed, Oct 26, 2011 at 3:30 AM, Mike Frysinger wrote:
>>>>> --- a/arch/sandbox/cpu/os.c
>>>>> +++ b/arch/sandbox/cpu/os.c
>>>>>
>>>>> +off_t os_lseek(int fd, off_t offset, int whence)
>>>>> +{
>>>>> +       return lseek(fd, offset, whence);
>>>>> +}
>>>>
>>>> How are we going to deal with calls to something like this where
>>>> 'whence' is defined in an OS-specific header file? I believe we need
>>>> to add definitions to os.h to provide this feature, and make them
>>>> match Posix. Is there a better way? open has a similar problem.
>>>
>>> yeah, i noticed this, but was ignoring it for the time being.  the
>>> only portable way i think is to have our local os.h provide arbitrary
>>> defines which the rest of u-boot will utilize, and then in the os_xxx
>>> func, we translate into the host defines.  a bit ugly, but it's the
>>> only way to really work since there is no guarantee that something
>>> like O_RDONLY will always be defined to 0.
>>
>> I care less about portability to non-POSIX, others may though. Perhaps
>> just defining the types and #defines with a suitable prefix in the
>> header file, and then a compile-time assert in the C would be good
>> enough.
>
> funny story: POSIX doesn't mandate the values behind the defines, only
> that the defines exist and are suitable for CPP usage.  notice i said
> "portable" and not "POSIX" ... that was not an accident :).  the
> defines may differ across Linux architectures, as well as ABIs (e.g.
> x86 vs x86_64 vs x32), as well as C libraries, as well as operating
> systems.  in practice, i vaguely recall the few defines we probably
> would use (ro/wo/rw/create) are the same across all Linux
> arches/ABIs/C libraries, but that's an icky situation to paint
> yourself into.

That's not very funny, but in any case it is manageable with a
sufficient amount of code bloat.

>
> a quick grep would probably clarify if the #if/#error assert logic
> would be sufficient to pass compile/run, but not necessarily pass the
> wd ;).
>
>>>>> --- a/arch/sandbox/cpu/start.c
>>>>> +++ b/arch/sandbox/cpu/start.c
>>>>>
>>>>> +       if (os_getopt("-h", 0) || os_getopt("--help", 0)) {
>>>>> +               serial_puts(usage);
>>>>> +               return 0;
>>>>> +       }
>>>>
>>>> I did actually create something to parse args and put them into a
>>>> 'state' structure. But it uses getopt_long(). Any reason we can't?
>>>
>>> on reason i didn't was so i wouldn't have to worry about translating
>>> defines/structs used in u-boot and defines/structs used by the host
>>> system.  i.e. the same issue we face with bits for os_{open,lseek}.
>>
>> Well I don't think we should be passing any getopt structures outside
>> os.c or parse.c. That is the intent of my state.c - a place where the
>> test state / arguments is stored and can be accessed when needed
>> without reference to os-dependent code / structures.
>
> hrm, i'm not seeing what you mean, but i'll take a closer look at the code.
> -mike
>

There isn't much to look at yet, sorry.

Regards,
Simon