Re: [PATCH] linux: Fallback to fork and exec for clone fail on posix_spawn (BZ#29115)

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 02/05/2022 17:08, Alexey Izbyshev wrote:
> On 2022-05-02 23:02, Adhemerval Zanella wrote:
>> On 02/05/2022 16:54, Alexey Izbyshev wrote:
>>> On 2022-05-02 22:42, Adhemerval Zanella wrote:
>>>> On 02/05/2022 16:15, Alexey Izbyshev wrote:
>>>>> On 2022-05-02 21:06, Adhemerval Zanella wrote:
>>>>>> +      while (__write_nocancel (args->pipe[1], &ret, sizeof (ret)) < 0);
>>>>>
>>>>> If the parent is killed, write() will fail with EPIPE, so this loop will become infinite. Maybe check for EPIPE and break?
>>>>
>>>> We block all signals (__libc_signal_block_all), so the parent will be
>>>> only killed
>>>> by signals that can no be caught (SIGKILL).  Same for SIGPIPE.
>>>
>>> The parent can indeed only be killed by SIGKILL, or simply exit from another thread, but how exactly it dies doesn't matter. But signals are already unblocked at this point in the child, so if SIGPIPE was ignored in the parent, it's ignored now in the child and is unblocked, hence write() can fail with EPIPE.
>>>
>>
>> Right, I was thinking in term of the CLONE_VFORK which is not the case.  But
>> now checking on the closefrom and the pipe file descriptor issue, it seems
>> it would be better to have the error code communicate through a shared memory
>> instead of a pipe.
> 
> How do you propose to wait until the child exec's if you use shared memory? Do you plan to use clone(CLONE_VFORK) (but without CLONE_VM)?

Your suggestion to use a two clone_range calls is way simpler than a shared map.