pinctrl: mvebu: armada-38x: add suspend/resume support

On Mon, Oct 19, 2015 at 11:25:49AM +0200, Marcin Wojtas wrote:
> Thomas,
> 
> 2015-10-19 9:23 GMT+02:00 Thomas Petazzoni
> <thomas.petazzoni@free-electrons.com>:
> > Hello,
> >
> > On Mon, 19 Oct 2015 08:04:49 +0200, Marcin Wojtas wrote:
> >
> >> > I don't like this. The mvebu_pinctrl_soc_info structure is meant to be
> >> > a read-only structure that only describes static information giving
> >> > SoC-specific details for pin-muxing. The idea is that in the event
> >> > where you had multiple pinctrl in the same system, you would still have
> >> > only one instance of mvebu_pinctrl_soc_info.
> >>
> >> Ok, understood. What with current static globals, like mpp_base? This
> >> is a problem when we consider hypothetical multi-pintrl system...
> >
> > The current driver is indeed not designed for multiple instances of the
> > same pinctrl controller. But that's exactly what Russell is asking for.
> >
> 
> In each mvebu pinctl there are <soc>_mpp_ctrl_get/set force usage of
> global mpp_base. In order not to use it this way the functions have to
> be redesigned. IMO having an acces to pdev would be sufficient (please
> see my proposal below), but yet I don't have a clear idea, how to do
> it,

I'm not sure having access to pdev would be sufficient -
struct mvebu_pinctrl is entirely private to pinctrl-mvebu.c, so you
wouldn't be able to dereference the driver data.

> > So it should be the other way around: the SoC-specific driver create a
> > structure, and this structure points back to the mvebu_pinctrl
> > structure.
> >
> 
> How about a following compromise:
> Instead of forcing pm_info, we can add a generic pointer to struct
> mvebu_pinctrl, e.g. void *priv; and enable passing it in a way I
> proposed before.
> 
> This way each SoC-specific driver can attach whatever is needed for
> its operation. For example in A38X/AXP it would be *base,
> *mpp_saved_regs and nregs gathered in one structure, but it wouldn't
> force anything specific for other machines. As a result in A38X/AXP in
> suspend/resume routines, there would be no globals in use at all.

The one thing I don't like about passing a 'void *' around is that it
will try to be re-used with __iomem pointers, and that'll cause sparse
to complain (quite rightfully - since sparse is designed to detect
direct dereferences of iomem pointers, and to do that it doesn't let
you trivially cast to/from an __iomem pointer to a normal pointer.
(not without using __force, which shouldn't be used in drivers.)

I wonder if there's a fairly nice solution to this:

Move away from mvebu_pinctrl_probe(), to
	void *mvebu_pinctrl_alloc(struct device *, size_t private_size);
	int mvebu_pinctrl_add(void *);

where mvebu_pinctrl_alloc() allocates and initialises the struct
mvebu_pinctrl, plus padding plus the private size.  The padding is
necessary to ensure that anything placed in the private area is
properly aligned (eg, 8-byte aligned.)  (This idea comes from netdev's
private data handling.)

These set the driver data to the private area, and return a pointer
to that private area.

Arrange for the get/set/whatever else functions to be passed this
same pointer.

What this means is that the individual drivers:

* get their private data wrapped up into the driver private data
* still don't need to know anything about mvebu_pinctrl
* can get at their private data using platform_get_drvdata()
* don't need any globals for anything

Something like this (by way of illustration):

 drivers/pinctrl/mvebu/pinctrl-armada-38x.c | 27 +++++++--
 drivers/pinctrl/mvebu/pinctrl-mvebu.c      | 97 ++++++++++++++++++++----------
 drivers/pinctrl/mvebu/pinctrl-mvebu.h      |  3 +
 3 files changed, 90 insertions(+), 37 deletions(-)

pinctrl: mvebu: armada-38x: add suspend/resume support

Commit Message

Patch