Message-ID: <20110525062312.GE22096@pengutronix.de>
Date:	Wed, 25 May 2011 08:23:12 +0200
From:	Sascha Hauer <s.hauer@...gutronix.de>
To:	Colin Cross <ccross@...gle.com>
Cc:	Jeremy Kerr <jeremy.kerr@...onical.com>, linux-sh@...r.kernel.org,
	Thomas Gleixner <tglx@...utronix.de>,
	"linux-arm-kernel@...ts.infradead.org" 
	<linux-arm-kernel@...ts.infradead.org>,
	lkml <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH 0/4] Add a generic struct clk

On Tue, May 24, 2011 at 12:41:20PM -0700, Colin Cross wrote:
> On Tue, May 24, 2011 at 1:09 AM, Sascha Hauer <s.hauer@...gutronix.de> wrote:
> > On Tue, May 24, 2011 at 12:31:13AM -0700, Colin Cross wrote:
> >> On Mon, May 23, 2011 at 11:26 PM, Sascha Hauer <s.hauer@...gutronix.de> wrote:
> >> > On Mon, May 23, 2011 at 04:12:24PM -0700, Colin Cross wrote:
> >> >> >
> >> >> >   tglx's plan is to create a separate struct clk_hwdata, which contains a
> >> >> >   union of base data structures for common clocks: div, mux, gate, etc. The
> >> >> >   ops callbacks are passed a clk_hw, plus a clk_hwdata, and most of the base
> >> >> >   hwdata fields are handled within the core clock code. This means less
> >> >> >   encapsulation of clock implementation logic, but more coverage of
> >> >> >   clock basics through the core code.
> >> >>
> >> >> I don't think they should be a union, I think there should be 3
> >> >> separate private datas, and three sets of clock ops, for the three
> >> >> different types of clock blocks: rate changers (dividers and plls),
> >> >> muxes, and gates.  These blocks are very often combined - a device
> >> >> clock often has a mux and a divider, and clk_set_parent and
> >> >> clk_set_rate on the same struct clk both need to work.
> >> >
> >> > The idea is to being able to propagate functions to the parent. It's
> >> > very convenient for the implementation of clocks when they only
> >> > implement either a divider, a mux or a gate. Combining all of these
> >> > into a single clock leads to complicated clock trees and many different
> >> > clocks where you can't factor out the common stuff.
> >>
> >> That seems complicated.  You end up with lots of extra clocks (meaning
> >> more boilerplate in the platform files) that have no meaning in the
> >> system (what is the clock between the mux and the divider in Tegra's
> >> i2c1 clock, it can never feed any devices), and you have to figure out
> >> at the framework level when to propagate and when to error out.  I'm
> >> not even sure you can always find the right place to stop propagating
> >> - do you just keep going up until the set_parent callback succeeds, or
> >> exists, or what?
> >
> > For the set_parent there would be no propagating at all. For set_rate I
> > can imagine a flag in the generic clock which tells whether to propagate
> > set_rate or not.
> >
> >>
> >> I think you can still factor out all the common code if you treat each
> >> clock as a possible mux, divider, and gate combination.  Each part of
> >> the clock is still just as abstractable as before - you can set the
> >> rate_ops to be the generic single register divider implementation, and
> >> the gate ops to be the generic single bit enable implementation.  The
> >> idea of what a clock node is matches the HW design,
> >
> > The hardware design consists of only discrete rate changers (plls,
> > dividers), muxes and gates. These are the only building blocks
> > *every* hardware design has. I believe that many of the problems
> > the current implementations have are due to the multiple building
> > blocks stuffed into one clock. If you haven't already take a look
> > at my i.MX5 clock patches:
> >
> > http://thread.gmane.org/gmane.linux.ports.arm.kernel/113631
> >
> > They need changes to fit onto the current patches and the rate
> > propagation problem is not solved there, but the resulting clock
> > data files are really short and nice to read. Furthermore it's easy
> > to implement. Just look at the diagrams in the datasheet and go through
> > them.
> 
> Propagation is what I'm trying simplify, because it's impossible at
> the framework level to determine the right time to propagate, and the
> right time to return an error, and I don't like pushing it down to
> each clock implementation either, because then you need a "propagating
> clk gate" and a "non-propagating clk gate".

Or a single flag on either the gate/mux/div or in the clock framework.

> 
> Your building blocks implement every op - clk_gate_ops implements
> set_rate as clk_parent_set_rate (won't that cause a deadlock on
> prepare_lock when it calls clk_set_rate?).  I'm suggesting breaking
> out the clock gate ops into a struct that only contains:
> enable
> disable
> prepare
> unprepare
> And a rate ops struct that contains:
> round_rate
> set_rate

+ get_rate?

> And a mux ops struct that contains
> set_parent
> 
> For a single HW clock that has a mux, a gate, and a divider, the
> existing implementation requires:
> INIT_CLK(..., clk_mux_ops)
> INIT_CLK(..., clk_div_ops)
> INIT_CLK(..., clk_gate_ops)
> which creates 3 clocks, and requires lots of propagation logic to
> figure out how to call clk_set_rate on the single clock that is
> exposed to the device, but not propagate it past the device clock if
> the device clock doesn't have a divider (propagating it up to an
> active pll feeding other devices results in disaster, at least on
> Tegra).  This combination doesn't seem to be common in your MX code,
> but _every_ Tegra device clock has these three parts.

Lucky you. Maybe the Tegra hardware designers have better drugs or just
take less of them...

> 
> With multiple smaller building blocks that can fit inside a clock, all
> you need is:
> INIT_CLK(..., clk_mux_ops, clk_div_ops, clk_gate_ops)
> You have one struct clk, which is exposed to the device and matches
> the datasheet.  If the clock has rate ops, clk_set_rate works with no
> propagation.

No, we have cascaded dividers and thus still need rate propagation.
Also we have the situation where a device is not allowed to adjust the
divider because other clocks are derived from the same divider. Still
the driver needs the rate. Here's an example from the i.MX51
(simplified):

-> PLL ----> divider -----> gate ---> MMC1
                     \
                      \
                       ---> gate ---> MMC2

Neither MMC1 nor MMC2 can be allowed to change the rate because it would
change the rate of the other MMC. Both MMCs need the rate to adjust the
card to the correct speed. In this case the platform code needs a
possibility to adjust the rate initially to a value suitable for both
MMC cores (50MHz). For this is needs access to the divider, preferably
as a clock to be able to use standard clock functions.

There are also many many dividers and muxes which are standalone
somewhere in the clock tree, not directly associated with any device.
With a clock which can be a mux and a divider together one is tempted
to combine these into the least possible number of clocks leading to
some random combination of clocks.


> If it doesn't have a rate, clk_rate_ops is NULL, and the
> framework deal with propagation only in the case where it is really
> propagation - a child clock that requires changing a parent clock.
> The block abstraction is still in place, there are just 3 slots for
> blocks within each clock.
> 
> Using a flag to mark clocks as "non-propagatable" is also not correct
> - propagatability is a feature of the parent, not the child.

this flag could also be added to the parent if we wish to.

Sascha

-- 
Pengutronix e.K.                           |                             |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists