Re: [take24 0/6] kevent: Generic event handling mechanism.

[Andrew Morton <akpm@osdl.org>]

On Mon, 20 Nov 2006 11:25:01 +0300
Evgeniy Polyakov <johnpol@2ka.mipt.ru> wrote:

> On Sun, Nov 19, 2006 at 04:02:03PM -0800, Ulrich Drepper (drepper@redhat.com) wrote:
> > Evgeniy Polyakov wrote:
> > >>Possible solution:
> > >>
> > >>a) it would be possible to have a "used" flag in each ring buffer entry.
> > >>   That's too expensive, I guess.
> > >>
> > >>b) kevent_wait needs another parameter which specifies the which is the
> > >>   last (i.e., least recently added) entry in the ring buffer.
> > >>   Everything between this entry and the current head (in ->kidx) is
> > >>   occupied.  If multiple threads arrive in kevent_wait the highest idx
> > >>   (with wrap around possibly lowest) is used.
> > >>
> > >>   kevent_wait will not try to move more entries into the ring buffer
> > >>   if ->kidx and the higest index passed in to any kevent_wait call
> > >>   is equal (i.e., the ring buffer is full).
> > >>
> > >>   There is one issue, though, and that is that a system call is needed
> > >>   to signal to the kernel that more entries in the ring buffer are
> > >>   processed and that they can be refilled.  This goes against the
> > >>   kernel filling the ring buffer automatically (see below)
> > >
> > >If thread calls kevent_wait() it means it has processed previous entries, 
> > >one can call kevent_wait() with $num parameter as zero, which
> > >means that thread does not want any new events, so nothing will be
> > >copied.
> > 
> > This doesn't solve the problem.  You could only request new events when 
> > all previously reported events are processed.  Plus: how do you report 
> > events if the you don't allow get_event pass them on?
> 
> Userspace should itself maintain order and possibility to get event in
> this implementation, kernel just returns events which were requested.

That would mean that in a multithreaded application (or multi-processes
sharing the same MAP_SHARED ringbuffer), all threads/processes will be
slowed down to wait for the slowest one.

> > >They all already imeplemented. Just all above, and it was done several
> > >months ago already. No need to reinvent what is already there.
> > >Even if we will decide to remove kevent_get_events() in favour of ring
> > >buffer-only implementation, winting-for-event syscall will be
> > >essentially kevent_get_events() without pointer to the place where to
> > >put events.
> > 
> > Right, but this limitation of the interface is important.  It means the 
> > interface of the kernel is smaller: fewer possibilities for problems and 
> > fewer constraints if in future something should be changed (and smaller 
> > kernel).
> 
> Ok, lets see for ring buffer implementation right now, and then we will
> decide if we want to remove or to stay with kevent_get_events() syscall.

I agree that kevent_get_events() is duplicative and we shouldn't need it. 
Better to concentrate all our development effort on the single and most
flexible means of delivery.