On 10/16/2015 05:57 PM, Jan Kiszka wrote:
> On 2015-10-16 17:31, Philippe Gerum wrote:
>> On 10/16/2015 05:28 PM, Jan Kiszka wrote:
>>> On 2015-10-16 17:22, Philippe Gerum wrote:
>>>> On 10/16/2015 04:56 PM, Jan Kiszka wrote:
>>>>> On 2015-10-16 16:49, Jan Kiszka wrote:
>>>>>> Hi,
>>>>>>
>>>>>> kill() is currently handled by libcobalt such that PIDs <= 0 are
>>>>>> forwarded to Linux and PIDs > 0 are considered to target only Xenomai
>>>>>> threads. But what if the user wants to address a regular Linux task from
>>>>>> within a Xenomai application? Shouldn't we retry kill via the Linux path
>>>>>> if Xenomai's syscall reports ESRCH?
>>>>>>
>>>>>
>>>>> IOW:
>>>>>
>>>>> diff --git a/lib/cobalt/signal.c b/lib/cobalt/signal.c
>>>>> index aac4059..7e03301 100644
>>>>> --- a/lib/cobalt/signal.c
>>>>> +++ b/lib/cobalt/signal.c
>>>>> @@ -99,6 +99,10 @@ COBALT_IMPL(int, kill, (pid_t pid, int sig))
>>>>>
>>>>> ret = XENOMAI_SYSCALL2(sc_cobalt_kill, pid, sig);
>>>>> if (ret) {
>>>>> + /* Retry with regular kill is no RT target was found. */
>>>>> + if (ret == -ESRCH)
>>>>> + return __STD(kill(pid, sig));
>>>>> +
>>>>> errno = -ret;
>>>>> return -1;
>>>>> }
>>>>>
>>>>> Jan
>>>>>
>>>>
>>>> This may break code that sends signal 0 to detect whether a rt thread
>>>> exists (like copperplate does), which is the reason for the lack of
>>>> forwarding IIRC. (ret == -ESRCH && sig) would be required to forward
>>>> without breaking such assumption.
>>>
>>> That still breaks POSIX (what if the user wants to test for a non-rt
>>> thread, like this is possible under regular Linux?). Can't copperplate
>>> be changed to bypass the wrapper?
>>>
>>
>> Probably, yes.
>>
>
> Looking at cluster_probe, I wonder if there is actually a problem.
> Doesn't that code also run over mercury? Then kill(pid, 0) also targets
> the whole system, not just a set of Xenomai applications.
Over mercury, the whole system is the rt domain.
Or do we need
> to ensure that the caller is not migrated to Linux needlessly under
> cobalt? IOW: can that probing happen under RT constraints?
>
It does with clusters. In general, no restriction on the calling domain
should exist for this low level interface.
> The same applies for me to the other users of the kill-based probing
> pattern (copperplate/heapobj-pshared.c and copperplate/regd/fs-common.c).
>
> Jan
>
--
Philippe.
_______________________________________________
Xenomai mailing list
Xenomai@xenomai.org
http://xenomai.org/mailman/listinfo/xenomai
