Re: [patch, rfc] lt-epoll ( level triggered epoll ) ...

[Davide Libenzi]

On Tue, 11 Mar 2003, bert hubert wrote:

> On Mon, Mar 10, 2003 at 12:15:25PM -0800, Davide Libenzi wrote:
>
> > 2) Existing apps using poll/select can easily be ported usinf LT epoll
>
> This is a big thing. I created a webserver based on MTasker
> (ds9a.nl/mtasker) that used select, poll or epoll and it was very hard to
> abstract this properly as level and edge semantics differ so wildly.
>
> Most programs will not abandon 'legacy' interfaces like poll and select and
> will only want to offer epoll in addition. Right now that is hard to do.

I agree here. It took 15 minutes to port thttpd to LT epoll.

> > 1) We leave epoll as is ( ET )
> > 2) We apply the patch that will make epoll LT
> > 3) We add a parameter to epoll_create() to fix the interface behaviour at
> > 	creation time ( small change on the current patch )
> >
> > With 2) and 3) there are also man pages to be reviewed to be posted to
> > Andries. Comments ?
>
> I'd vote for 2.

I received a bunch of private emails ( ppl that are using ET epoll )
asking me to have both behaviours. The code require no more than 10 lines
of code to give both possibilities. We have two options in doing that :

1)
We add a parameter to epoll_create() that will set the interface behaviour
at creation time :

	#define EPOLL_ET (1 << 0)

	int epoll_create(int size, int flags);

2)
We can go at fd granularity by leaving the API the same, and we define :

	#define EPOLLET (1 << 31)

	...
	events = EPOLLIN | EPOLLET;

What do you think ?

- Davide