Re: [PATCH 3/3] serial: 8250: Add a wakeup_capable module param

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Wednesday, January 18, 2012, Paul E. McKenney wrote:
> On Wed, Jan 18, 2012 at 02:15:59PM -0800, Simon Glass wrote:
> > Hi Paul,
> > 
> > On Wed, Jan 18, 2012 at 1:42 PM, Paul E. McKenney
> > <paulmck@linux.vnet.ibm.com> wrote:
> > > On Wed, Jan 18, 2012 at 01:08:13PM -0800, Simon Glass wrote:
> > >> [+cc Rafael J. Wysocki <rjw@sisk.pl> who I think wrote the wakeup.c code]
> > >>
> > >> Hi Alan, Paul,
> > >>
> > >> On Tue, Jan 17, 2012 at 8:17 PM, Paul E. McKenney
> > >> <paulmck@linux.vnet.ibm.com> wrote:
> > >> > On Tue, Jan 17, 2012 at 08:10:36PM +0000, Alan Cox wrote:
> > >> >> On Tue, 17 Jan 2012 10:56:03 -0800
> > >> >> Simon Glass <sjg@chromium.org> wrote:
> > >> >>
> > >> >> > Since serial_core now does not make serial ports wake-up capable by
> > >> >> > default, add a parameter to support this feature in the 8250 UART.
> > >> >> > This is the only UART where I think this feature is useful.
> > >> >>
> > >> >> NAK
> > >> >>
> > >> >> Things should just work for users. Magic parameters is not an
> > >> >> improvement. If its a performance problem someone needs to fix the rcu
> > >> >> sync overhead or stop using rcu on that path.
> > >>
> > >> OK fair enough, I agree. Every level I move down the source tree
> > >> affects more people though.
> > >>
> > >> >
> > >> > I must say that I lack context here, even after looking at the patch,
> > >> > but the synchronize_rcu_expedited() primitives can be used if the latency
> > >> > of synchronize_rcu() is too large.
> > >> >
> > >>
> > >> Let me provide a bit of context. The serial_core code seems to be the
> > >> only place in the kernel that does this:
> > >>
> > >>               device_init_wakeup(tty_dev, 1);
> > >>               device_set_wakeup_enable(tty_dev, 0);
> > >>
> > >> The first call makes the device wakeup capable and enables wakeup, The
> > >> second call disabled wakeup.
> > >>
> > >> The code that removes the wakeup source looks like this:
> > >>
> > >> void wakeup_source_remove(struct wakeup_source *ws)
> > >> {
> > >>       if (WARN_ON(!ws))
> > >>               return;
> > >>
> > >>       spin_lock_irq(&events_lock);
> > >>       list_del_rcu(&ws->entry);
> > >>       spin_unlock_irq(&events_lock);
> > >>       synchronize_rcu();
> > >> }
> > >>
> > >> The sync is there because we are about to destroy the actual ws
> > >> structure (in wakeup_source_destroy()). I wonder if it should be in
> > >> wakeup_source_destroy() but that wouldn't help me anyway.
> > >>
> > >> synchronize_rcu_expedited() is a bit faster but not really fast
> > >> enough. Anyway surely people will complain if I put this in the wakeup
> > >> code - it will affect all wakeup users. It seems to me that the right
> > >> solution is to avoid enabling and then immediately disabling wakeup.
> > >
> > > Hmmm...  What hardware are you running this one?  Normally,
> > > synchronize_rcu_expedited() will be a couple of orders of magnitude
> > > faster than synchronize_rcu().
> > >
> > >> I assume we can't and shouldn't change device_init_wakeup() . We could
> > >> add a call like device_init_wakeup_disabled() which makes the device
> > >> wakeup capable but does not actually enable it. Does that work?
> > >
> > > If the only reason for the synchronize_rcu() is to defer the pair of
> > > kfree()s in wakeup_source_destroy(), then another possible approach
> > > would be to remove the synchronize_rcu() from wakeup_source_remove()
> > > and then use call_rcu() to defer the two kfree()s.
> > >
> > > If this is a reasonable change to make, the approach is as follows:
> > >
> > > 1.      Add a struct rcu_head to wakeup_source, call it "rcu".
> > >        Or adjust the following to suit your choice of name.
> > >
> > > 2.      Replace the pair of kfree()s with:
> > >
> > >                call_rcu(&ws->rcu, wakeup_source_destroy_rcu);
> > >
> > > 3.      Create the wakeup_source_destroy_rcu() as follows:
> > >
> > >        static void wakeup_source_destroy_rcu(struct rcu_head *head)
> > >        {
> > >                struct wakeup_source *ws =
> > >                        container_of(head, struct wakeup_source, rcu);
> > >
> > >                kfree(ws->name);
> > >                kfree(ws);
> > >        }
> > >
> > > Of course, this assumes that it is OK for wakeup_source_unregister()
> > > to return before the memory is freed up.  This often is OK, but there
> > > are some cases where the caller requires that there be no further
> > > RCU readers with access to the old data.  In these cases, you really
> > > do need the wait.
> > 
> > Thanks very much for that. I'm not sure if it is a reasonable change,
> > but it does bug me that we add it to a data structure knowing that we
> > will immediately remove it!
> > 
> > >From what I can see, making a device wakeup-enabled mostly happens on
> > init or in response to a request to the driver (presumably from user
> > space). In the latter case I suspect the synchronise_rcu() is fine. In
> > the former it feels like we should make up our minds which of the
> > three options is required (incapable, capable but not enabled, capable
> > and enabled).
> > 
> > I will try a patch first based on splitting the two options (capable
> > and enable) and see if that get a NAK.
> > 
> > Then I will come back to your solution - it seems fine to me and not a
> > lot of code. Do we have to worry about someone enabling, disabled,
> > enabling and then disabling wakeup quickly? Will this method break in
> > that case if the second call to call_rcu() uses the same wc->rcu?
> 
> There are a couple of questions here, let me take them one at a time:
> 
> 1.	If you just disabled, can you immediately re-enable?
> 
> 	The answer is "yes".  The reason that this works is that you
> 	allocate a new structure for the re-enabling, and that new
> 	structure has its own rcu_head field.
> 
> 2.	If you repeatedly disable and re-enable in a tight loop,
> 	can this cause problems?
> 
> 	The answer to this is also "yes" -- you can run the system
> 	out of memory doing that.  However, there are a number of
> 	simple ways to avoid this problem:
> 
> 	a.	Do a synchronize_rcu() on every (say) thousandth
> 		disable operation.
> 
> 	b.	As above, but only do the synchronize_rcu() if
> 		all 1,000 disable operations occurred within
> 		(say) a second of each other.
> 
> 	c.	As above, but actually count the number of
> 		pending call_rcu() callbacks.
> 
> 	Both (a) and (b) can be carried out on a per-CPU basis if there
> 	is no convenient locked structure in which to track the state.
> 	You cannot carry (c) out on a per-CPU basis because RCU callbacks
> 	can sometimes be invoked on a different CPU from the one that
> 	call_rcu()ed them.  Rare, but it can happen.
> 
> 	I would expect that option (a) would work in almost all cases.
> 
> If this can be exercised freely from user space, then you probably
> really do need #2 above.

Yes, you can, but then I'd say it's not necessary for user space to
be able to carry that out in a tight loop.  So, it seems, alternatively,
we could make that loop a bit less tight, e.g. by adding an arbitrary
sleep to the user space interface for the "disable" case.

Thanks,
Rafael