Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Conor Dooley]

[-- Attachment #1: Type: text/plain, Size: 6279 bytes --]

Stephen,

On Thu, Oct 03, 2024 at 12:33:40PM +0100, Conor Dooley wrote:
> On Wed, Oct 02, 2024 at 03:21:16PM +0200, Jerome Brunet wrote:
> > On Wed 02 Oct 2024 at 13:20, Neil Armstrong <neil.armstrong@linaro.org> wrote:
> > 
> > > On 02/10/2024 12:48, Conor Dooley wrote:
> > >> From: Conor Dooley <conor.dooley@microchip.com>
> > >> I like this one better than qualcomms and wish to use it for the
> > >> PolarFire SoC clock drivers.
> > >> Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> > >> ---
> > >>   drivers/clk/Kconfig                           |  4 ++
> > >>   drivers/clk/Makefile                          |  1 +
> > >>   drivers/clk/{meson => }/clk-regmap.c          |  2 +-
> > >>   drivers/clk/meson/Kconfig                     | 46 +++++++++----------
> > >>   drivers/clk/meson/Makefile                    |  1 -
> > >>   drivers/clk/meson/a1-peripherals.c            |  2 +-
> > >>   drivers/clk/meson/a1-pll.c                    |  2 +-
> > >>   drivers/clk/meson/axg-aoclk.c                 |  2 +-
> > >>   drivers/clk/meson/axg-audio.c                 |  2 +-
> > >>   drivers/clk/meson/axg.c                       |  2 +-
> > >>   drivers/clk/meson/c3-peripherals.c            |  2 +-
> > >>   drivers/clk/meson/c3-pll.c                    |  2 +-
> > >>   drivers/clk/meson/clk-cpu-dyndiv.c            |  2 +-
> > >>   drivers/clk/meson/clk-dualdiv.c               |  2 +-
> > >>   drivers/clk/meson/clk-mpll.c                  |  2 +-
> > >>   drivers/clk/meson/clk-phase.c                 |  2 +-
> > >>   drivers/clk/meson/clk-pll.c                   |  2 +-
> > >>   drivers/clk/meson/g12a-aoclk.c                |  2 +-
> > >>   drivers/clk/meson/g12a.c                      |  2 +-
> > >>   drivers/clk/meson/gxbb-aoclk.c                |  2 +-
> > >>   drivers/clk/meson/gxbb.c                      |  2 +-
> > >>   drivers/clk/meson/meson-aoclk.h               |  2 +-
> > >>   drivers/clk/meson/meson-eeclk.c               |  2 +-
> > >>   drivers/clk/meson/meson-eeclk.h               |  2 +-
> > >>   drivers/clk/meson/meson8-ddr.c                |  2 +-
> > >>   drivers/clk/meson/meson8b.c                   |  2 +-
> > >>   drivers/clk/meson/s4-peripherals.c            |  2 +-
> > >>   drivers/clk/meson/s4-pll.c                    |  2 +-
> > >>   drivers/clk/meson/sclk-div.c                  |  2 +-
> > >>   drivers/clk/meson/vclk.h                      |  2 +-
> > >>   drivers/clk/meson/vid-pll-div.c               |  2 +-
> > >>   .../meson => include/linux/clk}/clk-regmap.h  |  0
> > >>   32 files changed, 53 insertions(+), 53 deletions(-)
> > >>   rename drivers/clk/{meson => }/clk-regmap.c (99%)
> > >>   rename {drivers/clk/meson => include/linux/clk}/clk-regmap.h (100%)
> > >> 
> > > <snip>
> > >
> > > I don't have objections, but I think Stephen didn't like the idea
> > > a few years ago, but perhaps it has changed...
> > >
> > > Anyway, take my:
> > > Acked-by: Neil Armstrong <neil.armstrong@linaro.org>
> > 
> > We had a similar discussion 3y ago indeed:
> > https://lore.kernel.org/linux-clk/162734682512.2368309.12015873010777083014@swboyd.mtv.corp.google.com/
> > 
> > There are needs for a common regmap backed clocks indeed, but allowing
> > meson flavored regmap clocks to spread in the kernel was not really the
> > prefered way to do it. 
> 
> Cool, thanks for that link.
> 
> > IIRC, Stephen's idea was more the bring regmap support in clk-gate.c,
> > clk-mux, etc ... I'm not quite sure how make iomem and regmap co-exist
> > in a manageable/maintainable way within those drivers (without adding yet
> > another level of abstraction I mean) ? Silently creating a regmap maybe
> > ? but that's probably a bit heavy. I did not really had time to dig more
> > on this, I guess no one did.
> 
> I guess I have some motivation to looking into it at the moment. I had
> my reservations about the Meson approach too, liking it more than
> Qualcomm's didn't mean I completely liked it.
> It was already my intention to implement point b of your mail, had the
> general idea here been acceptable, cos that's a divergence from how the
> generic clock types (that the driver in question currently uses) work.
> And on that note, I just noticed I left the mild-annoyance variable name
> "sigh" in the submitted driver changes, which I had used for the
> clk_regmap struct that your point b in the link relates to.
> 
> > I don't really have a preference one way or the other but if it is going
> > to be exposed in 'include/linux', we need to be sure that's how we want
> > to do it. With clocks poping in many driver subsystems, it will
> > difficult to change afterward. 
> 
> Yeah, I agree. I didn't expect this to go in right away, and I also
> didn't want to surge ahead on some rework of the clock types, were
> people to hate even the reuse.

Hmm, so how (in-)complete of a regmap implementation can I get away
with? I only need clk-gate and clk-divider for this patchset...

Shoving the regmap into the clk structs makes things pretty trivial as I
don't need to do anything special in any function other than
clk_*_readl()/clk_*_writel() and the registration code. A flag isn't
even needed to determine if a clock is a regmap one I don't think, since
you can just check if the regmap pointer is non-NULL. My use case doesn't
actually need the registration code changes either as, currently, only reg
gets set at runtime, but leaving that out is a level of incomplete I'd not
let myself away with.
Obviously shoving the extra members into the clk structs has the downside
of taking up a pointer and a offset worth of memory for each clock of
that type registered, but it is substantially easier to support devices
with multiple regmaps that way. Probably moot though since the approach you
suggested in the thread linked above that implements a clk_hw_get_regmap()
has to store a pointer to the regmap's identifier which would take up an
identical amount of memory.

I don't really care which way you want it done, both are pretty easy to
implement if I can get away with just doing so for the two standard
clock types that I am using - is it okay to just do those two?

Cheers,
Conor.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Stephen Boyd]

Quoting Conor Dooley (2024-11-06 04:56:25)
> On Thu, Oct 03, 2024 at 12:33:40PM +0100, Conor Dooley wrote:
> > > IIRC, Stephen's idea was more the bring regmap support in clk-gate.c,
> > > clk-mux, etc ... I'm not quite sure how make iomem and regmap co-exist
> > > in a manageable/maintainable way within those drivers (without adding yet
> > > another level of abstraction I mean) ? Silently creating a regmap maybe
> > > ? but that's probably a bit heavy. I did not really had time to dig more
> > > on this, I guess no one did.
> > 
> > I guess I have some motivation to looking into it at the moment. I had
> > my reservations about the Meson approach too, liking it more than
> > Qualcomm's didn't mean I completely liked it.
> > It was already my intention to implement point b of your mail, had the
> > general idea here been acceptable, cos that's a divergence from how the
> > generic clock types (that the driver in question currently uses) work.
> > And on that note, I just noticed I left the mild-annoyance variable name
> > "sigh" in the submitted driver changes, which I had used for the
> > clk_regmap struct that your point b in the link relates to.
> > 
> > > I don't really have a preference one way or the other but if it is going
> > > to be exposed in 'include/linux', we need to be sure that's how we want
> > > to do it. With clocks poping in many driver subsystems, it will
> > > difficult to change afterward. 
> > 
> > Yeah, I agree. I didn't expect this to go in right away, and I also
> > didn't want to surge ahead on some rework of the clock types, were
> > people to hate even the reuse.
> 
> Hmm, so how (in-)complete of a regmap implementation can I get away
> with? I only need clk-gate and clk-divider for this patchset...
> 
> Shoving the regmap into the clk structs makes things pretty trivial as I
> don't need to do anything special in any function other than
> clk_*_readl()/clk_*_writel() and the registration code. A flag isn't
> even needed to determine if a clock is a regmap one I don't think, since
> you can just check if the regmap pointer is non-NULL.

For the basic clk types I think it would be good to leave the old stuff
alone. We have already split the logic out into helpers, so I wonder if
we can do this better by making kernel modules for the different basic
regmap clk types and exposing registration APIs. If we force drivers
that use the basic regmap types to 'select' the module then we'll make
it so that we don't include code that isn't used anywhere. That's one of
the problems with the basic clk types, it's always built. It also lets
us avoid making regmap a dependency for the clk framework at large.

Doing that would also let us avoid the flag because it will be explicit
in any registration API, clk_register_divider() vs.
clk_register_regmap_divider(). Yes we duplicate some boiler plate logic
around read-only and registration paths, but this is alright as long as
we can share most of the code and gain the advantage of removing the
code entirely when it isn't used.

I wonder if we can even make a regmap on the fly for the iomem pointers
so that clk_divider_readl() can always use the regmap API to access the
hardware. Sounds wasteful but maybe it would work.

> My use case doesn't
> actually need the registration code changes either as, currently, only reg
> gets set at runtime, but leaving that out is a level of incomplete I'd not
> let myself away with.
> Obviously shoving the extra members into the clk structs has the downside
> of taking up a pointer and a offset worth of memory for each clock of
> that type registered, but it is substantially easier to support devices
> with multiple regmaps that way. Probably moot though since the approach you
> suggested in the thread linked above that implements a clk_hw_get_regmap()
> has to store a pointer to the regmap's identifier which would take up an
> identical amount of memory.

We don't need to store the regmap identifier in the struct clk. We can
store it in the 'struct clk_init_data' with some new field, and only do
that when/if we actually need to. We would need to pass the init data to
the clk_ops::init() callback though. We currently knock that out during
registration so that clk_hw->init is NULL. Probably we can just set that
to NULL after the init routine runs in __clk_core_init().

Long story short, don't add something to 'struct clk_core', 'struct
clk', or 'struct clk_hw' for these details. We can have a 'struct
clk_regmap_hw' that everyone else can build upon:

  struct clk_regmap_hw {
        struct regmap *regmap;
        struct clk_hw hw;
  };

and then set the regmap pointer during registration in
clk_hw_init_regmap().

int clk_hw_init_regmap(struct clk_hw *hw)
{
	struct device *dev;
	struct regmap *regmap;
	struct clk_regmap_hw *rhw;

	rhw = clk_hw_to_clk_regmap_hw(hw);

	dev = clk_hw_get_dev(hw);
	if (!dev)
		return -EINVAL;

	regmap = dev_get_regmap(dev, hw->init->regmap_name);
	if (!regmap)
		return -EINVAL; // Print helpful message
	rhw->regmap = regmap;

	return 0;
}

or we can even make it so that there's clk_hw_init_regmap() and
clk_hw_init_regmap_name() so that drivers can have multiple functions if
the clks need different regmaps.

int my_init_regmap(struct clk_hw *hw)
{
	int ret;

	ret = clk_hw_init_regmap_name(hw, "my_name");
	...
}

If you don't need the multiple regmap support then it's fine to punt
here until later.

> 
> I don't really care which way you want it done, both are pretty easy to
> implement if I can get away with just doing so for the two standard
> clock types that I am using - is it okay to just do those two?
> 

Of course, doing only what is minimally required is better than changing
everything if you're not sure the approach is going to land.

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Conor Dooley]

[-- Attachment #1: Type: text/plain, Size: 2250 bytes --]

On Thu, Nov 14, 2024 at 05:29:54PM -0800, Stephen Boyd wrote:
> Quoting Conor Dooley (2024-11-06 04:56:25)
> > My use case doesn't
> > actually need the registration code changes either as, currently, only reg
> > gets set at runtime, but leaving that out is a level of incomplete I'd not
> > let myself away with.
> > Obviously shoving the extra members into the clk structs has the downside
> > of taking up a pointer and a offset worth of memory for each clock of
> > that type registered, but it is substantially easier to support devices
> > with multiple regmaps that way. Probably moot though since the approach you
> > suggested in the thread linked above that implements a clk_hw_get_regmap()
> > has to store a pointer to the regmap's identifier which would take up an
> > identical amount of memory.
> 
> We don't need to store the regmap identifier in the struct clk. We can
> store it in the 'struct clk_init_data' with some new field, and only do
> that when/if we actually need to. We would need to pass the init data to
> the clk_ops::init() callback though. We currently knock that out during
> registration so that clk_hw->init is NULL. Probably we can just set that
> to NULL after the init routine runs in __clk_core_init().
> 
> Long story short, don't add something to 'struct clk_core', 'struct
> clk', or 'struct clk_hw' for these details. We can have a 'struct
> clk_regmap_hw' that everyone else can build upon:
> 
>   struct clk_regmap_hw {
>         struct regmap *regmap;
>         struct clk_hw hw;
>   };

What's the point of this? I don't understand why you want to do this over
what clk_divider et al already do, where clk_hw and the iomem pointer
are in the struct itself.

> 
> and then set the regmap pointer during registration in
> clk_hw_init_regmap().
> 
> int clk_hw_init_regmap(struct clk_hw *hw)
> {
> 	struct device *dev;
> 	struct regmap *regmap;
> 	struct clk_regmap_hw *rhw;
> 
> 	rhw = clk_hw_to_clk_regmap_hw(hw);
> 
> 	dev = clk_hw_get_dev(hw);
> 	if (!dev)
> 		return -EINVAL;
> 
> 	regmap = dev_get_regmap(dev, hw->init->regmap_name);
> 	if (!regmap)
> 		return -EINVAL; // Print helpful message
> 	rhw->regmap = regmap;
> 
> 	return 0;
> }

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Stephen Boyd]

Quoting Conor Dooley (2024-11-28 02:36:16)
> On Thu, Nov 14, 2024 at 05:29:54PM -0800, Stephen Boyd wrote:
> > Quoting Conor Dooley (2024-11-06 04:56:25)
> > > My use case doesn't
> > > actually need the registration code changes either as, currently, only reg
> > > gets set at runtime, but leaving that out is a level of incomplete I'd not
> > > let myself away with.
> > > Obviously shoving the extra members into the clk structs has the downside
> > > of taking up a pointer and a offset worth of memory for each clock of
> > > that type registered, but it is substantially easier to support devices
> > > with multiple regmaps that way. Probably moot though since the approach you
> > > suggested in the thread linked above that implements a clk_hw_get_regmap()
> > > has to store a pointer to the regmap's identifier which would take up an
> > > identical amount of memory.
> > 
> > We don't need to store the regmap identifier in the struct clk. We can
> > store it in the 'struct clk_init_data' with some new field, and only do
> > that when/if we actually need to. We would need to pass the init data to
> > the clk_ops::init() callback though. We currently knock that out during
> > registration so that clk_hw->init is NULL. Probably we can just set that
> > to NULL after the init routine runs in __clk_core_init().
> > 
> > Long story short, don't add something to 'struct clk_core', 'struct
> > clk', or 'struct clk_hw' for these details. We can have a 'struct
> > clk_regmap_hw' that everyone else can build upon:
> > 
> >   struct clk_regmap_hw {
> >         struct regmap *regmap;
> >         struct clk_hw hw;
> >   };
> 
> What's the point of this? I don't understand why you want to do this over
> what clk_divider et al already do, where clk_hw and the iomem pointer
> are in the struct itself.

Can you give an example? I don't understand what you're suggesting. I
prefer a struct clk_regmap_hw like above so that the existing struct
clk_hw in the kernel aren't increased by a pointer. SoC drivers can use
the same struct as a replacement for their struct clk_hw member today.

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Conor Dooley]

[-- Attachment #1: Type: text/plain, Size: 3265 bytes --]

On Tue, Dec 03, 2024 at 02:50:31PM -0800, Stephen Boyd wrote:
> Quoting Conor Dooley (2024-11-28 02:36:16)
> > On Thu, Nov 14, 2024 at 05:29:54PM -0800, Stephen Boyd wrote:
> > > Quoting Conor Dooley (2024-11-06 04:56:25)
> > > > My use case doesn't
> > > > actually need the registration code changes either as, currently, only reg
> > > > gets set at runtime, but leaving that out is a level of incomplete I'd not
> > > > let myself away with.
> > > > Obviously shoving the extra members into the clk structs has the downside
> > > > of taking up a pointer and a offset worth of memory for each clock of
> > > > that type registered, but it is substantially easier to support devices
> > > > with multiple regmaps that way. Probably moot though since the approach you
> > > > suggested in the thread linked above that implements a clk_hw_get_regmap()
> > > > has to store a pointer to the regmap's identifier which would take up an
> > > > identical amount of memory.
> > > 
> > > We don't need to store the regmap identifier in the struct clk. We can
> > > store it in the 'struct clk_init_data' with some new field, and only do
> > > that when/if we actually need to. We would need to pass the init data to
> > > the clk_ops::init() callback though. We currently knock that out during
> > > registration so that clk_hw->init is NULL. Probably we can just set that
> > > to NULL after the init routine runs in __clk_core_init().
> > > 
> > > Long story short, don't add something to 'struct clk_core', 'struct
> > > clk', or 'struct clk_hw' for these details. We can have a 'struct
> > > clk_regmap_hw' that everyone else can build upon:
> > > 
> > >   struct clk_regmap_hw {
> > >         struct regmap *regmap;
> > >         struct clk_hw hw;
> > >   };
> > 
> > What's the point of this? I don't understand why you want to do this over
> > what clk_divider et al already do, where clk_hw and the iomem pointer
> > are in the struct itself.
> 
> Can you give an example? I don't understand what you're suggesting. I
> prefer a struct clk_regmap_hw like above so that the existing struct
> clk_hw in the kernel aren't increased by a pointer. SoC drivers can use
> the same struct as a replacement for their struct clk_hw member today.

Best example I guess is to link what I did? This one is the core
changes:
https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=35904222355e971c24b3eb9b9fad3dd0c38d1393
clk-gate has my original hack that I did while trying to figure out
what you wanted, clk-divider-regmap is a 99% copy of clk-divider with
the types, function names and readl()/writel() implementations modified.
Before your last set of comments I was doing something identical to the
clk-gate change for clk-divider also.
Here's the changes required to my driver to make it work with the
updated:
https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=ea40211fe20f8bc6ef0320b93e1baa5b3f244601
It's pretty much a drop in replacement, other than the additional
complexity in probe.

Hopefully that either gets my point across or lets you spot why I don't
understand the benefit of a wrapper around clk_hw.

Cheers,
Conor.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Conor Dooley]

[-- Attachment #1: Type: text/plain, Size: 3513 bytes --]

Hey Stephen,

Any thoughts on the example I gave below?

On Fri, Dec 06, 2024 at 01:56:08PM +0000, Conor Dooley wrote:
> On Tue, Dec 03, 2024 at 02:50:31PM -0800, Stephen Boyd wrote:
> > Quoting Conor Dooley (2024-11-28 02:36:16)
> > > On Thu, Nov 14, 2024 at 05:29:54PM -0800, Stephen Boyd wrote:
> > > > Quoting Conor Dooley (2024-11-06 04:56:25)
> > > > > My use case doesn't
> > > > > actually need the registration code changes either as, currently, only reg
> > > > > gets set at runtime, but leaving that out is a level of incomplete I'd not
> > > > > let myself away with.
> > > > > Obviously shoving the extra members into the clk structs has the downside
> > > > > of taking up a pointer and a offset worth of memory for each clock of
> > > > > that type registered, but it is substantially easier to support devices
> > > > > with multiple regmaps that way. Probably moot though since the approach you
> > > > > suggested in the thread linked above that implements a clk_hw_get_regmap()
> > > > > has to store a pointer to the regmap's identifier which would take up an
> > > > > identical amount of memory.
> > > > 
> > > > We don't need to store the regmap identifier in the struct clk. We can
> > > > store it in the 'struct clk_init_data' with some new field, and only do
> > > > that when/if we actually need to. We would need to pass the init data to
> > > > the clk_ops::init() callback though. We currently knock that out during
> > > > registration so that clk_hw->init is NULL. Probably we can just set that
> > > > to NULL after the init routine runs in __clk_core_init().
> > > > 
> > > > Long story short, don't add something to 'struct clk_core', 'struct
> > > > clk', or 'struct clk_hw' for these details. We can have a 'struct
> > > > clk_regmap_hw' that everyone else can build upon:
> > > > 
> > > >   struct clk_regmap_hw {
> > > >         struct regmap *regmap;
> > > >         struct clk_hw hw;
> > > >   };
> > > 
> > > What's the point of this? I don't understand why you want to do this over
> > > what clk_divider et al already do, where clk_hw and the iomem pointer
> > > are in the struct itself.
> > 
> > Can you give an example? I don't understand what you're suggesting. I
> > prefer a struct clk_regmap_hw like above so that the existing struct
> > clk_hw in the kernel aren't increased by a pointer. SoC drivers can use
> > the same struct as a replacement for their struct clk_hw member today.
> 
> Best example I guess is to link what I did? This one is the core
> changes:
> https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=35904222355e971c24b3eb9b9fad3dd0c38d1393
> clk-gate has my original hack that I did while trying to figure out
> what you wanted, clk-divider-regmap is a 99% copy of clk-divider with
> the types, function names and readl()/writel() implementations modified.
> Before your last set of comments I was doing something identical to the
> clk-gate change for clk-divider also.
> Here's the changes required to my driver to make it work with the
> updated:
> https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=ea40211fe20f8bc6ef0320b93e1baa5b3f244601
> It's pretty much a drop in replacement, other than the additional
> complexity in probe.
> 
> Hopefully that either gets my point across or lets you spot why I don't
> understand the benefit of a wrapper around clk_hw.
> 
> Cheers,
> Conor.



[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v1 08/11] clk: move meson clk-regmap implementation to common code

[Conor Dooley]

[-- Attachment #1: Type: text/plain, Size: 4217 bytes --]

On Tue, Jan 21, 2025 at 05:38:53PM +0000, Conor Dooley wrote:
> Hey Stephen,
> 
> Any thoughts on the example I gave below?

I'll give you a few more days to comment, and then I'll just send a
fresh revision, implemented as below. The links seem to have expired in
a rebase along the way, so I've provided some fresh links.

Cheers,
Conor.

> 
> On Fri, Dec 06, 2024 at 01:56:08PM +0000, Conor Dooley wrote:
> > On Tue, Dec 03, 2024 at 02:50:31PM -0800, Stephen Boyd wrote:
> > > Quoting Conor Dooley (2024-11-28 02:36:16)
> > > > On Thu, Nov 14, 2024 at 05:29:54PM -0800, Stephen Boyd wrote:
> > > > > Quoting Conor Dooley (2024-11-06 04:56:25)
> > > > > > My use case doesn't
> > > > > > actually need the registration code changes either as, currently, only reg
> > > > > > gets set at runtime, but leaving that out is a level of incomplete I'd not
> > > > > > let myself away with.
> > > > > > Obviously shoving the extra members into the clk structs has the downside
> > > > > > of taking up a pointer and a offset worth of memory for each clock of
> > > > > > that type registered, but it is substantially easier to support devices
> > > > > > with multiple regmaps that way. Probably moot though since the approach you
> > > > > > suggested in the thread linked above that implements a clk_hw_get_regmap()
> > > > > > has to store a pointer to the regmap's identifier which would take up an
> > > > > > identical amount of memory.
> > > > > 
> > > > > We don't need to store the regmap identifier in the struct clk. We can
> > > > > store it in the 'struct clk_init_data' with some new field, and only do
> > > > > that when/if we actually need to. We would need to pass the init data to
> > > > > the clk_ops::init() callback though. We currently knock that out during
> > > > > registration so that clk_hw->init is NULL. Probably we can just set that
> > > > > to NULL after the init routine runs in __clk_core_init().
> > > > > 
> > > > > Long story short, don't add something to 'struct clk_core', 'struct
> > > > > clk', or 'struct clk_hw' for these details. We can have a 'struct
> > > > > clk_regmap_hw' that everyone else can build upon:
> > > > > 
> > > > >   struct clk_regmap_hw {
> > > > >         struct regmap *regmap;
> > > > >         struct clk_hw hw;
> > > > >   };
> > > > 
> > > > What's the point of this? I don't understand why you want to do this over
> > > > what clk_divider et al already do, where clk_hw and the iomem pointer
> > > > are in the struct itself.
> > > 
> > > Can you give an example? I don't understand what you're suggesting. I
> > > prefer a struct clk_regmap_hw like above so that the existing struct
> > > clk_hw in the kernel aren't increased by a pointer. SoC drivers can use
> > > the same struct as a replacement for their struct clk_hw member today.
> > 
> > Best example I guess is to link what I did? This one is the core
> > changes:
> > https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=35904222355e971c24b3eb9b9fad3dd0c38d1393

https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=435c8eb223ee804297a0491fae2b00d3d5a9c773

> > clk-gate has my original hack that I did while trying to figure out
> > what you wanted, clk-divider-regmap is a 99% copy of clk-divider with
> > the types, function names and readl()/writel() implementations modified.
> > Before your last set of comments I was doing something identical to the
> > clk-gate change for clk-divider also.
> > Here's the changes required to my driver to make it work with the
> > updated:
> > https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=ea40211fe20f8bc6ef0320b93e1baa5b3f244601

https://git.kernel.org/pub/scm/linux/kernel/git/conor/linux.git/commit/?h=syscon-rework-2&id=f55e907e93c55c943725dd62c2fc7dc76cdbd8d5

> > It's pretty much a drop in replacement, other than the additional
> > complexity in probe.
> > 
> > Hopefully that either gets my point across or lets you spot why I don't
> > understand the benefit of a wrapper around clk_hw.
> > 
> > Cheers,
> > Conor.
> 
> 



[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]