Re: [PATCH] signal: Fix the error return of kill -1

["Eric W. Biederman" <ebiederm@xmission.com>]

"Eric W. Biederman" <ebiederm@xmission.com> writes:

> Oleg Nesterov <oleg@redhat.com> writes:
>
>> On 08/16, Eric W. Biederman wrote:
>>>
>>> Oleg Nesterov <oleg@redhat.com> writes:
>>>
>>> > On 08/15, David Laight wrote:
>>> >>
>>> >> or maybe even:
>>> >> 	} else {
>>> >> 		struct task_struct * p;
>>> >> 		int err;
>>> >> 		ret = -ESRCH;
>>> >>
>>> >> 		for_each_process(p) {
>>> >> 			if (task_pid_vnr(p) > 1 &&
>>> >> 					!same_thread_group(p, current)) {
>>> >> 				err = group_send_sig_info(sig, info, p,
>>> >> 							  PIDTYPE_MAX);
>>> >> 				if (ret)
>>> >> 					ret = err;
>>> >
>>> > Hmm, indeed ;)
>>> >
>>> > and "err" can be declared inside the loop.
>>>
>>> We can't remove the success case, from my posted patch.
>>>
>>> A signal is considered as successfully delivered if at least
>>> one process receives it.
>>
>> Yes.
>>
>> Initially ret = -ESRCH.
>>
>> Once group_send_sig_info() succeeds at least once (returns zero)
>> ret becomes 0.
>>
>> After that
>>
>> 	if (ret)
>> 		ret = err;
>>
>> has no effect.
>>
>> So if a signal is successfully delivered at least once the code
>> above returns zero.
>
> Point.
>
> We should be consistent and ensure  __kill_pgrp_info uses
> the same code pattern, otherwise it will be difficult to
> see they use the same logic.
>
> Does "if (ret) ret = err;" generate better code than "success |= !err"?
>

I just looked at the assembly output and at least on x86 with cmov
"if (ret) ret = err;" generates the better assembly even in
the inner loop.

> I think for both patterns the reader of the code is going to have to
> stop and think about what is going on to understand the logic.
>
> We should probably do something like:
>
> 	/* 0 for success or the last error */
> 	if (ret)
>         	ret = err;
>

Even with that comment it feels awkward to me.

Does anyone have any idea how to make that idiom more obvious
what is happening?

Eric