[En-Nut-Discussion] Event API

[Matthias Ringwald]

Hi there,

On 27.10.2004, at 10:02, Harald Kipp wrote:

>
>>
>> NutEventInit/Reset/Clear:
>>
>> NutEnterCritical()
>> theEvent = 0
>> NutExitCritical()
>
> Matthias' post let me think about NutEventBroadcast().
> Do we need a critical section in the SIGNALED case?

I'd say, there should be a critical section no matter if SIGNALED or 
not.

> I assume, we don't, but I'm not sure.
>
> int NutEventBroadcast(HANDLE * qhp)
> {
>     int rc = 0;
>
>     if (*qhp == SIGNALED) { <---- UNPROTECTED
>         *qhp = 0;           <---- UNPROTECTED
>     else
>         while (*qhp)
>             rc += NutEventPost(qhp);
>
>     return rc;
> }

Examples (although pathological)

A: qhq == 0

Thread A enters NutEventBroadcast
>     if (*qhp == SIGNALED) { <---- UNPROTECTED
reads one byte of *qhp (it's a 16 bit pointer)
interrupt is triggered
interrupt call NutEventPostAsync
*qhp is set in such a way that the second byte of *qhp matches SIGNALED 
(afaik 0xffff)
here one could argue, that Ethernut per default has no mem at 0xffxx 
and structs
might not be on odd addresses. so this should work in practice, at 
least in theory it's unclean.

B: qhp == SIGNALED

Thread A enters NutEventBroadcast
>     if (*qhp == SIGNALED) { <---- UNPROTECTED
>         *qhp = 0;           <---- UNPROTECTED
after first byte write interrupt is triggered so
*qhp = 0xffXX of 0xXXff
interrupt call NutEventPostAsync and *qph != SIGNALED, so thread *qhp 
is woken
=> CRASH


wrapping up: yes, for SIGNALED case critical section is a must, for the 
other it would be better/cleaner.
btw. we use 64k RAM, so that a thread struct at 0xffXX is not 
impossible.

harald, will you add a critical section here?
otherwise, I have to!

cheers,
matthias