Re: [PATCH] Revised timerfd() interface

[Davide Libenzi]

On Thu, 6 Sep 2007, Michael Kerrisk wrote:

> Hi Davide,
> 
> > > > > > > As I think about this more, I see more problems with
> > > > > > > your argument.  timerfd needs the ability to get and 
> > > > > > > get-while-setting just as much as the earlier APIs.
> > > > > > > Consider a library that creates a timerfd file descriptor that
> > > > > > > is handed off to an application: that library may want
> > > > > > > to modify the timer settings without having to create a
> > > > > > > new file descriptor (the app mey not be able to be told about
> > > > > > > the new fd).  Your argument just doesn't hold, AFAICS.
> > > > > > 
> > > > > > Such hypotethical library, in case it really wanted to offer such 
> > > > > > functionality, could simply return an handle instead of the raw
> > > > > > fd, and take care of all that stuff in userspace.
> > > > > 
> > > > > Did I miss something?  Is it not the case that as soon as the
> > > > > library returns a handle, rather than an fd, then the whole
> > > > > advantage of timerfd() (being able to select/poll/epoll on 
> > > > > the timer as well as other fds) is lost?  
> > > > 
> > > > Why? The handle would simply be a little struct where the timerfd fd
> > > > is 
> > > > stored, and a XXX_getfd() would return it.
> > > > So my point is, I doubt such functionalities are really needed, and I 
> > > > also argue that the kernel is the best place for such wrapper code
> > > > to go.
> > > 
> > > So what happens if one thread (via the library) wants modify
> > > a timer's settings at the same timer as another thread is 
> > > select()ing on it?  The first thread can't do this by creating
> > > a new timerfd timer, since it wants to affect the select()
> > > in the other thread?
> > 
> > It can be done w/out any problems. The select thread will be notified 
> > whenever the new timer setting expires.
> 
> We are going in circles here.  I think you are missing my point.
> Consider the following
> 
> [[
> Thread A: calls library function which creates a timerfd file
> descriptor.
> 
> Thread B: calls select() on the timerfd file descriptor.
> 
> Thread A: calls library function which wants to:
>    a) modify timer settings, and retrieve copy of current timer
>       settings, and later
>    b) restore old timer settings.
> ]]
> 
> This seems a quite reasonable use-case to me, and the existing
> interface simply can't support it.

"Quite reasonable"? :)
I honestly doubt it, but anyway. Modulo error checking:

struct tfd {
	int fd, clockid;
	struct itimerspec ts;
};

struct tfd *tfd_create(int clockid, int flags, const struct itimerspec *ts) {
	struct tfd *th;
	th = malloc(sizeof(*th));
	th->clockid = clockid;
	th->ts = *ts;
	th->fd = timerfd(-1, clockid, flags, ts);
	return th;
}

void tfd_close(struct tfd *th) {
	close(th->fd);
	free(th);
}

int tfd_getfd(const struct tfd *th) {
	return th->fd;
}

int tfd_gettime(const struct tfd *th, int *clockid, struct itimerspec *ts) {
	*clockid = th->clockid;
	*ts = th->ts;
	return 0;
}

int tfd_settime(struct tfd *th, int clockid, int flags,
		const struct itimerspec *ts) {
	th->fd = timerfd(th->fd, clockid, flags, ts);
	th->clockid = clockid;
	th->ts = *ts;
	return 0;
}

Wrap the get/set with a mutex in case you plan to shoot yourself in a foot 
by doing get/set from multiple threads ;)
So, once again:

- I sincerly doubt the above is common usage/design patters for timerfds

  * timerfds are not a common global resource, ala signals, that requires 
    get+set+restore pattern - you can have many of them set to different 
    times

- Those IMO *very* special use cases can be handled in userspace with few 
  lines of code, *if* really needed



- Davide


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/