[U-Boot] [PATCH v3 00/11] clk: Port Linux common clock framework [CCF] to U-boot (tag: 5.0-rc3)

[Simon Glass]

Hi Lukasz,

On Sun, 19 May 2019 at 15:03, Lukasz Majewski <lukma at denx.de> wrote:
>
> Hi Simon,
>
> > Hi Lukasz,
> >
> > On Sat, 18 May 2019 at 15:28, Lukasz Majewski <lukma at denx.de> wrote:
> > >
> > > Hi Simon,
> > >
> > > This is not the newest patch set version of CCF (v3 vs. v4), but the
> > > comments/issues apply.
> > >
> > > > Hi Lukasz,
> > > >
> > > > On Thu, 25 Apr 2019 at 04:30, Lukasz Majewski <lukma at denx.de>
> > > > wrote:
> > > > >
> > > > > This patch series brings the files from Linux kernel to provide
> > > > > clocks support as it is used on the Linux kernel with common
> > > > > clock framework [CCF] setup.
> > > > >
> > > > > This series also fixes several problems with current clocks and
> > > > > provides sandbox tests for functions addded to clk-uclass.c
> > > > > file.
> > > > >
> > > > > Design decisions/issues:
> > > > > =========================
> > > > >
> > > > > - U-boot's DM for clk differs from Linux CCF. The most notably
> > > > > difference is the lack of support for hierarchical clocks and
> > > > > "clock as a manager driver" (single clock DTS node acts as a
> > > > > starting point for all other clocks).
> > > > >
> > > > > - The clk_get_rate() now caches the previously read data (no
> > > > > need for recursive access.
> > > > >
> > > > > - On purpose the "manager" clk driver (clk-imx6q.c) is not using
> > > > > large table to store pointers to clocks - e.g.
> > > > > clk[IMX6QDL_CLK_USDHC2_SEL] = .... Instead we use udevice's
> > > > > linked list for the same class (UCLASS_CLK). The rationale -
> > > > > when porting the code as is from Linux, one would need ~1KiB of
> > > > > RAM to store it. This is way too much if we do plan to use this
> > > > > driver in SPL.
> > > > >
> > > > > - The "central" structure of this patch series is struct udevice
> > > > > and its driver_data field contains the struct clk pointer (to
> > > > > the originally created one).
> > > > >
> > > > > - Up till now U-boot's driver model's CLK operates on udevice
> > > > > (main access to clock is by udevice ops)
> > > > >   In the CCF the access to struct clk (comprising pointer to
> > > > > *dev) is possible via dev_get_driver_data()
> > > > >
> > > > >   Storing back pointer (from udevice to struct clk) as
> > > > > driver_data is a convention for CCF.
> > > >
> > > > Ick. Why not use uclass-private data to store this, since every
> > > > UCLASS_CLK device can have a parent.
> > >
> > > The "private_data" field would be also a good place to store the
> > > back pointer from udevice to struct clk [*]. The problem with CCF
> > > and udevice's priv pointer is explained just below:
> > >
> > > >
> > > > >
> > > > > - I could use *private_alloc_size to allocate driver's 'private"
> > > > >   structures (dev->priv) for e.g. divider (struct clk_divider
> > > > > *divider) for IMX6Q clock, but this would change the original
> > > > > structure of the CCF code.
> > >
> > > The original Linux's CCF code for iMX relies on using kmalloc
> > > internally:
> > >
> > > https://elixir.bootlin.com/linux/v5.1.2/source/drivers/clk/imx/clk-gate2.c#L139
> > > https://elixir.bootlin.com/linux/v5.1.2/source/drivers/clk/clk-divider.c#L471
> > >
> > > By using driver_data I've avoided the need to make more changes to
> > > the original Linux code.
> > >
> > > I could use udevice's priv with automatic data allocation but then
> > > the CCF ported code would require more changes and considering the
> > > (from the outset) need to "fit" this code into U-Boot's DM, it
> > > drives away from the original Linux code.
> >
> > Is the main change the need to cast driver_data?
>
> The main change would be to remove the per clock device memory
> allocation code (with exit paths) from the original CCF code.
>
> This shall not be so difficult.
>
> > Perhaps that could be
> > hidden in a helper function/macro, so that in U-Boot it can hide the
> > use of (struct clk_uc_priv *)dev_get_uclass_priv(clk->dev))>parent  ?
>
> Helper function would help to some extend as we operate on structures
> similar to:
>
> struct clk_gate2 {
>         struct clk clk;
>         void __iomem    *reg;
>         u8              bit_idx;
>         u8              cgr_val;
>         u8              flags;
> };
>
> The helper would return struct clk's address which is the same as
> struct's clk_gate2 (this is assured by C standard).
>
> >
> > >
> > >
> > > > >
> > > > > The question is if it would be better to use private_alloc_size
> > > > > (and dev->private) or stay with driver_data.
> > > > > The former requires some rewritting in CCF original code (to
> > > > > remove (c)malloc, etc), but comply with u-boot DM. The latter
> > > > > allows re-using the CCF code as is, but introduces some
> > > > > convention special for CCF (I'm not sure thought if dev->priv
> > > > > is NOT another convention as well).
> > > >
> > > > Yes I would like to avoid malloc() calls in drivers and use the
> > > > in-built mechanism.
> > >
> > > I see your point.
> > >
> > > If the community agrees - I can rewrite the code to use such
> > > approach (but issues pointed out in [*] still apply).
> > >
> > > >
> > > > >
> > > > > - I've added the clk_get_parent(), which reads parent's
> > > > > dev->driver_data to provide parent's struct clk pointer. This
> > > > > seems the easiest way to get child/parent relationship for
> > > > > struct clk in U-boot's udevice based clocks.
> > > > >
> > > > > - For tests I had to "emulate" CCF code structure to test
> > > > > functionality of clk_get_parent_rate() and clk_get_by_id().
> > > > > Those functions will not work properly with "standard" (i.e.
> > > > > non CCF) clock setup(with not set dev->driver_data to struct
> > > > > clk).
> > > >
> > > > Well I think we need a better approach for that anywat.
> > > > driver_data is used for getting something from the DT.
> > >
> > > Maybe the name (driver_data) was a bit misleading then. For CCF it
> > > stores the back pointer to struct clk (as in fact it is a CCF's
> > > "driver data").
> >
> > Well it seems like a hack to me. Perhaps there is a good reason for it
> > in Linux?
>
> In Linux there is another approach - namely the struct clk (which is
> the main struct for clock management) has pointer to struct clk_core,
> which has pointer to parent(s).
>
> https://elixir.bootlin.com/linux/v5.1.2/source/drivers/clk/clk.c#L43
>
> In the case of U-Boot - the CCF wants to work on struct clk, but the
> _main_ data structure for U-Boot is struct udevice. Hence the need to
> have a back pointer (or force struct clk to have NOT pointer to udevice,
> but the udevice itself - then container_of would then do the trick).

The thing I don't understand is that I assumed there is no 1:1
correspondence from struct clk to struct udevice. I thought that we
could have one clock device which supports lots of clk IDs (e.g.
0-23).

>
> > Or is it just convenience?
>
> As stated above - Linux all necessary information has accessible from
> struct clk.

Sure, but we can always find the udevice from the clk.

If we require that clk == udevice then we can go back the other way
too, by using uclass-private data attached to each device.

>
> >
> > >
> > >
> > >
> > > NOTE:
> > >
> > > [*] - I do have a hard time to understand how struct clk shall work
> > > with struct udevice?
> > >
> > > In Linux or Barebox the struct clk is the "main" structure to hold
> > > the clock management data (like freq, ops, flags, parent/sibling
> > > relation, etc).
> >
> > Yes U-Boot has a uniform struct udevice for every device and struct
> > uclass for every class.
> >
> > But the interesting thing here is that clocks have their own
> > parent/sibling relationships, quite independent from the device tree.
>
> But there would be no harm if we could re-use udevice for it. In the
> current CCF (v4) patch set each clk IP block (like mux or gate) is
> modelled as udevice:
>
> https://pastebin.com/uVmwv5FT

I don't see how you can do this...doesn't it mean changing the parents
of existing devices? E.g. if a SPI clock can come from one of 4
parents, do you need to changes its parent in the driver-model tree?

>
> >
> > >
> > > A side observation - we now have three different implementations of
> > > struct clk in U-Boot :-) (two of which have *ops inside :-) )
> >
> > Oh dear.
> >
> > The broadcom iMX ones needs to be converted.
> >
> > >
> > > In the case of U-Boot's DM (./include/clk.h) it only has a
> > > _pointer_ to udevice (which means that I cannot get the struct clk
> > > easily from udevice with container_of()). The struct udevice has
> > > instead the *ops and *parent pointer (to another udevice).
> >
> > Yes that's correct. The struct clk is actually a handle to the clock,
> > and includes an ID number.
>
> You mean the ID number of the clock ?

Yes:

struct clk {
struct udevice *dev;
/*
* Written by of_xlate. We assume a single id is enough for now. In the
* future, we might add more fields here.
*/
unsigned long id;
};


>
> >
> > >
> > > Two problems:
> > >
> > > - Linux CCF code uses massively "struct clk" to handle clock
> > > operations (but not udevice)
> >
> > OK.
> >
> > >
> > > - There is no clear idea of how to implement
> > > struct clk *clk_get_parent(struct clk *) in U-Boot.
> >
> > As above, it seems that this might need to be implemented. I don't
> > think the DM parent/child relationships are good enough for clk, since
> > they are not aware of the ID.
> >
> > >
> > > The reason is that we lack clear information about which udevice's
> > > data field shall be used to store the back pointer from udevice to
> > > struct clk.
> > >
> > > Any hints and ideas are more than welcome.
> >
> > I think it would be good to get Stephen Warren's thoughts on this as
> > he made the change to introduce struct clk.
>
> Ok.
>
> >
> > But at present clk_set_parent() is implemented by calling into the
> > driver. The uclass itself does not maintain information about what is
> > a parent of what.
>
> Linux struct clk has easy access to its parent's struct clk. This is
> what is missing in U-Boot's DM.
>
> >
> > Do we *need* to maintain this information in the uclass?
>
> I think that we don't need. It would be enough to modify struct clk to
> has struct udevice embedded in it (as Linux has struct clk_core), not
> the pointer. Then we can use container_of to get clock and re-use
> struct udevice's parent pointer (and maybe UCLASS_CLK list of devices).
>
> >
> > I think it would be prohibitively expensive to separate out each
> > individual clock into a separate device (udevice), but that would
> > work.
>
> This is the approach as I use now in CCF v4:
> https://pastebin.com/uVmwv5FT
>
> It is expensive but logically correct as each mux, gate, pll is the
> other CLK IP block (device).

OK I see. What is the cost? Is it acceptable for a boot loader?

>
> >
> > The only other option I see is to create a sibling list and parent
> > pointer inside struct clk.
>
> This would be the approach similar to Linux kernel approach.
>
> However, I don't know what were original needs of struct clk (as it did
> not have it). Maybe Stephen can shed some light on it?

Hopefully.

>
> >
> > I suspect this will affect power domain also, although we don't have
> > that yet.
>
> iMX8 has some clocks which needs to be always recalculated as they
> depend on power domains which can be disabled.

OK

>
> >
> > Do you think there is a case for building this into DM itself, such
> > that devices can have a list of IDs for each device, each with
> > independent parent/child relationships?
>
> For iMX6Q the ID of the clock is used to get proper clock in drivers
> (or from DTS). All clock's udevices are stored into UCLASS_CLK list.
> With the ID we get proper udevice and from it and via driver_data the
> struct clk, which is then used in the CCF code to operate on clock
> devices (PLL, gate, mux, etc).
>
> I simply re-used the DM facilities (only missing is the back pointer
> from udevice to struct clk).

Well then you can use dev_get_uclass_priv() for that.

Regards,
Simon