Re: [RESEND][RFC PATCH v2] waitfd

[Casey Dahlin <cdahlin@redhat.com>]

Oleg Nesterov wrote:
> On 01/10, Casey Dahlin wrote:
>   
>> Scott James Remnant wrote:
>>     
>>> On Wed, 2009-01-07 at 18:19 +0100, Oleg Nesterov wrote:
>>>
>>>       
>>>> Please note that unlike other sys_...fd() syscalls, sys_waitfd()
>>>> doesn't allow to pass O_CLOEXEC. Looks like we need a separate
>>>> "flags" argument...
>>>>
>>>> Also, ioctl(FIONBIO) or fcntl(O_NONBLOCK) have no effect on
>>>> waitfd, not very good.
>>>>
>>>> I'd suggest to remove WNOHANG from waitfd_ctx->ops and treat
>>>> (->f_flags & O_NONBLOCK) as WNOHANG.
>>>>
>>>> (can't resist, ->ops is not the best name ;)
>>>>
>>>>
>>>>         
>>> Definitely agree here, waitfd() doesn't need WNOHANG - we already have
>>> ONONBLOCK.
>>>
>>> That also solves one of the strangest behaves of waitid when you use
>>> WNOHANG (it returns zero and you have to check whether it changed the
>>> struct), now you just read() - if no child you get EAGAIN, if a child
>>> you read a struct.
>>>
>>>       
>> From the perspective of waitfd, the only difference between WNOHANG and
>> O_NONBLOCK is which argument you put the flags in.
>>     
>
> No. Please see the note about ioctl/fcntl above.
>
> Oleg.
>   
Yes but the actual waitfd call could simply set O_NONBLOCK on the 
descriptor when it receive WNOHANG in the flags, and read the descriptor 
flags going forward.

--CJD