Jan Kiszka wrote:
> Philippe Gerum wrote:
>> We should indeed postpone this just in case the upper layer indexes the extra
>> state on the minor value. We can also simplify a few things doing so.
>>
>> --- ksrc/nucleus/pipe.c (revision 4565)
>> +++ ksrc/nucleus/pipe.c (working copy)
>> @@ -77,11 +77,9 @@
>>
>> static inline void xnpipe_minor_free(int minor)
>> {
>> - if (minor < 0 || minor >= XNPIPE_NDEVS)
>> - return;
>> -
>> - __clrbits(xnpipe_bitmap[minor / BITS_PER_LONG],
>> - 1UL << (minor % BITS_PER_LONG));
>> + /* May be called with nklock free. */
>> + clrbits(xnpipe_bitmap[minor / BITS_PER_LONG],
>> + 1UL << (minor % BITS_PER_LONG));
>
> Bad news: This doesn't fly as is. All modifying operations on
> xnpipe_bitmap must be atomic and xnpipe_bitmap has to be
> xnarch_atomic_t. But then find_first_zero_bit breaks. Is there some
> version for atomic arrays? I guess we have to open-code this, at least
> down to word-level...
>
#define clrbits(flags,mask) xnarch_atomic_clear_mask(&(flags),mask)
which is either translated to atomic_clear_mask() on x86 which works on a long
value, or to any Xenomai-local implementation that work the same way for other
archs. IOW, clrbits() shall be an atomic op by essence in our implementation.
The thing is that Linux atomic_set/clear_mask() on x86_64 is broken
(s,andl,andq), so we may have to provide ours.
> Jan
>
--
Philippe.
_______________________________________________
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core
