Re: [PATCH] [RFC] notify userspace about time changes

[Alexander Shishkin <virtuoso@slind.org>]

On Thu, Aug 05, 2010 at 03:38:54 -0700, Greg KH wrote:
> On Fri, Aug 06, 2010 at 12:22:27AM +0200, Kay Sievers wrote:
> > On Thu, Aug 5, 2010 at 23:38, Greg KH <gregkh@suse.de> wrote:
> > > On Thu, Aug 05, 2010 at 02:11:05PM -0700, john stultz wrote:
> > >> On Thu, 2010-08-05 at 15:33 +0300, Alexander Shishkin wrote:
> > >> > On 4 August 2010 18:58, john stultz <johnstul@us.ibm.com> wrote:
> > >> > > Is there a actual use case that you need this for?  I don't really have
> > >> > > an issue with the code I just really want to make sure the feature would
> > >> > > be useful enough to justify the API and code maintenance going forward.
> > >> >
> > >> > Yes. What we have here is an application which takes care of different means
> > >> > of time synchronization (trusted time servers, different GSM operators, etc)
> > >> > and also different kinds of time-based events/notifications (like "dentist
> > >> > appointment next thursday"). When it encounters a time change that is
> > >> > made by some other application, it basically wants to disable automatic
> > >> > time adjustment and trigger the events/notifications which are due at this
> > >> > (new) time.
> > >>
> > >> Ok. Something specific is always more helpful then theoretical uses.
> > >>
> > >> I think the filtering is still a bit controversial, so you might want to
> > >> respin it without that. But otherwise I'm ok with it as long as no one
> > >> else objects to any of the minor details of the interface
> > >>
> > >> GregKH: Does /sys/kernel/time_notify seem ok by you?
> > >
> > > Um, it depends, what is that file going to do?  I don't see a
> > > Documentation/ABI/ entry here that describes it fully :)
> > 
> > I think that's really awkward interface, to pass file descriptor
> > numbers around and write them to magic sysfs files.
> 
> Ick, really?  That's not ok for sysfs.

This is how cgroups notifications work [1], for example. One obtains an
eventfd, one then writes its descriptor to a certain control file, one polls
said eventfd for events. The type of event is echoed to the same control file
together with the fd.

So, my idea was, basically to use the sysfs file as means to specify what
kind of time-related events a program wants to subscribe to. For example,

<int:eventfd descriptor> <bool:want other's time changes> <bool:want own time
changes> <bool:want settimeofday> <bool:want adjtime>

> > I would very much prefer a file that contains the current time, and
> > wakes up possible users with a POLL_ERR on changes caused by some
> > other process. That works very well for things like /proc/mounts, is
> > easy to get, and does not need a full page of weird instructions to
> > get stuff done. :)
> 
> That sounds more reasonable.

Polling a simple sysfs file is not enough as I described in the other mail,
unless it implements a counter similar to eventfd, which would effectively
make it another eventfd implementation, only with a dentry this time.

[1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob;f=kernel/cgroup.c;h=a8ce099544049e787464c3e4261ac70625cc0653;hb=HEAD#l3071 onwards

Regards,
--
Alex