Stephen Boyd <[email protected]> writes:
> Quoting Maxime Ripard (2022-11-09 03:00:45)
>> On Mon, Nov 07, 2022 at 08:57:22PM +0000, Aidan MacDonald wrote:
>> >
>> > Maxime Ripard <[email protected]> writes:
>> >
>> > > Hi,
>> > >
>> > > On Fri, Nov 04, 2022 at 05:35:29PM +0000, Aidan MacDonald wrote:
>> >
>> > Assigning the parent clock in the DT works once, at boot, but going off
>> > what you wrote in the commit message, if the clock driver has a
>> > .determine_rate() implementation that *can* reparent clocks then it
>> > probably *will* reparent them, and the DT assignment will be lost.
>>
>> Yes, indeed, but assigned-clock-parents never provided any sort of
>> guarantee on whether or not the clock was allowed to reparent or not.
>> It's just a one-off thing, right before probe, and a clk_set_parent()
>> call at probe will override that just fine.
>>
>> Just like assigned-clock-rates isn't permanent.
>>
>> > What I'm suggesting is a runtime constraint that the clock subsystem
>> > would enforce, and actively prevent drivers from changing the parent.
>> > Either explicitly with clk_set_parent() or due to .determine_rate().
>> >
>> > That way you could write a .determine_rate() implementation that *can*
>> > select a better parent, but if the DT applies a constraint to fix the
>> > clock to a particular parent, the clock subsystem will force that parent
>> > to be used so you can be sure the clock is never reparented by accident.
>>
>> Yeah, that sounds like a good idea, and CLK_SET_RATE_NO_REPARENT isn't
>> too far off from this, it's just ignored by clk_set_parent() for now. I
>> guess we could rename CLK_SET_RATE_NO_REPARENT to CLK_NO_REPARENT, make
>> clk_set_parent handle it, and set that flag whenever
>> assigned-clock-parents is set on a clock.
>>
>> It's out of scope for this series though, and I certainly don't want to
>> deal with all the regressions it might create :)
>>
>
> This sounds like a new dt binding that says the assigned parent should
> never change. It sounds sort of like gpio hogs. A clock-hogs binding?
Ideally we want the clock driver to be able to reparent clocks freely
to get the best rate. But we also need some control over that to stop
consumers from being reparented in undesired ways. Eg. you might want
to make sure the GPU gets its own PLL so it can be reclocked easily,
and putting another device on the GPU's PLL could prevent that.
The only way to achieve this today is (1) never do any reparenting in
the clock driver; and (2) use assigned-clock-parents in the DT to set
up the entire clock tree manually.
Maxime said that (2) is basically wrong -- if assigned-clock-parents
provides no guarantee on what the OS does "after boot" then the OS is
pretty much free to ignore it.
My suggestion: add a per-clock bitmap to keep track of which parents
are allowed. Any operation that would select a parent clock not on the
whitelist should fail. Automatic reparenting should only select from
clocks on the whitelist. And we need new DT bindings for controlling
the whitelist, for example:
clock-parents-0 = <&clk1>, <&pll_c>;
clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
This means that clk1 can only have pll_c as a parent, while clk2 can
have pll_a or pll_b as parents. By default every clock will be able
to use any parent, so a list is only needed if the machine needs a
more restrictive policy.
assigned-clock-parents should disable automatic reparenting, but allow
explicit clk_set_parent(). This will allow clock drivers to start doing
reparenting without breaking old DTs.
Hi,
On Thu, Mar 23, 2023 at 03:35:30PM +0000, Aidan MacDonald wrote:
>
> Stephen Boyd <[email protected]> writes:
>
> > Quoting Maxime Ripard (2022-11-09 03:00:45)
> >> On Mon, Nov 07, 2022 at 08:57:22PM +0000, Aidan MacDonald wrote:
> >> >
> >> > Maxime Ripard <[email protected]> writes:
> >> >
> >> > > Hi,
> >> > >
> >> > > On Fri, Nov 04, 2022 at 05:35:29PM +0000, Aidan MacDonald wrote:
> >> >
> >> > Assigning the parent clock in the DT works once, at boot, but going off
> >> > what you wrote in the commit message, if the clock driver has a
> >> > .determine_rate() implementation that *can* reparent clocks then it
> >> > probably *will* reparent them, and the DT assignment will be lost.
> >>
> >> Yes, indeed, but assigned-clock-parents never provided any sort of
> >> guarantee on whether or not the clock was allowed to reparent or not.
> >> It's just a one-off thing, right before probe, and a clk_set_parent()
> >> call at probe will override that just fine.
> >>
> >> Just like assigned-clock-rates isn't permanent.
> >>
> >> > What I'm suggesting is a runtime constraint that the clock subsystem
> >> > would enforce, and actively prevent drivers from changing the parent.
> >> > Either explicitly with clk_set_parent() or due to .determine_rate().
> >> >
> >> > That way you could write a .determine_rate() implementation that *can*
> >> > select a better parent, but if the DT applies a constraint to fix the
> >> > clock to a particular parent, the clock subsystem will force that parent
> >> > to be used so you can be sure the clock is never reparented by accident.
> >>
> >> Yeah, that sounds like a good idea, and CLK_SET_RATE_NO_REPARENT isn't
> >> too far off from this, it's just ignored by clk_set_parent() for now. I
> >> guess we could rename CLK_SET_RATE_NO_REPARENT to CLK_NO_REPARENT, make
> >> clk_set_parent handle it, and set that flag whenever
> >> assigned-clock-parents is set on a clock.
> >>
> >> It's out of scope for this series though, and I certainly don't want to
> >> deal with all the regressions it might create :)
> >>
> >
> > This sounds like a new dt binding that says the assigned parent should
> > never change. It sounds sort of like gpio hogs. A clock-hogs binding?
>
> Ideally we want the clock driver to be able to reparent clocks freely
> to get the best rate. But we also need some control over that to stop
> consumers from being reparented in undesired ways. Eg. you might want
> to make sure the GPU gets its own PLL so it can be reclocked easily,
> and putting another device on the GPU's PLL could prevent that.
>
> The only way to achieve this today is (1) never do any reparenting in
> the clock driver; and (2) use assigned-clock-parents in the DT to set
> up the entire clock tree manually.
>
> Maxime said that (2) is basically wrong -- if assigned-clock-parents
> provides no guarantee on what the OS does "after boot" then the OS is
> pretty much free to ignore it.
I didn't really say it's wrong, just that it never provided the
guarantee you expect it to provide. I can't really say whether it's an
issue or not on your platform.
It's mostly unrelated to this series though, none of these patches
affect that behavior in one way or the other.
> My suggestion: add a per-clock bitmap to keep track of which parents
> are allowed. Any operation that would select a parent clock not on the
> whitelist should fail. Automatic reparenting should only select from
> clocks on the whitelist. And we need new DT bindings for controlling
> the whitelist, for example:
>
> clock-parents-0 = <&clk1>, <&pll_c>;
> clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
>
> This means that clk1 can only have pll_c as a parent, while clk2 can
> have pll_a or pll_b as parents. By default every clock will be able
> to use any parent, so a list is only needed if the machine needs a
> more restrictive policy.
>
> assigned-clock-parents should disable automatic reparenting, but allow
> explicit clk_set_parent(). This will allow clock drivers to start doing
> reparenting without breaking old DTs.
I'm generally not a fan of putting all these policies in the device
tree. Do you have an example where it wouldn't be possible to do exactly
this from the driver itself?
Maxime
Maxime Ripard <[email protected]> writes:
> On Thu, Mar 23, 2023 at 03:35:30PM +0000, Aidan MacDonald wrote:
>>
>> Stephen Boyd <[email protected]> writes:
>>
>> > Quoting Maxime Ripard (2022-11-09 03:00:45)
>> >> On Mon, Nov 07, 2022 at 08:57:22PM +0000, Aidan MacDonald wrote:
>> >> >
>> >> > Maxime Ripard <[email protected]> writes:
>> >> >
>> >> > > Hi,
>> >> > >
>> >> > > On Fri, Nov 04, 2022 at 05:35:29PM +0000, Aidan MacDonald wrote:
>> >> >
>> >> > Assigning the parent clock in the DT works once, at boot, but going off
>> >> > what you wrote in the commit message, if the clock driver has a
>> >> > .determine_rate() implementation that *can* reparent clocks then it
>> >> > probably *will* reparent them, and the DT assignment will be lost.
>> >>
>> >> Yes, indeed, but assigned-clock-parents never provided any sort of
>> >> guarantee on whether or not the clock was allowed to reparent or not.
>> >> It's just a one-off thing, right before probe, and a clk_set_parent()
>> >> call at probe will override that just fine.
>> >>
>> >> Just like assigned-clock-rates isn't permanent.
>> >>
>> >> > What I'm suggesting is a runtime constraint that the clock subsystem
>> >> > would enforce, and actively prevent drivers from changing the parent.
>> >> > Either explicitly with clk_set_parent() or due to .determine_rate().
>> >> >
>> >> > That way you could write a .determine_rate() implementation that *can*
>> >> > select a better parent, but if the DT applies a constraint to fix the
>> >> > clock to a particular parent, the clock subsystem will force that parent
>> >> > to be used so you can be sure the clock is never reparented by accident.
>> >>
>> >> Yeah, that sounds like a good idea, and CLK_SET_RATE_NO_REPARENT isn't
>> >> too far off from this, it's just ignored by clk_set_parent() for now. I
>> >> guess we could rename CLK_SET_RATE_NO_REPARENT to CLK_NO_REPARENT, make
>> >> clk_set_parent handle it, and set that flag whenever
>> >> assigned-clock-parents is set on a clock.
>> >>
>> >> It's out of scope for this series though, and I certainly don't want to
>> >> deal with all the regressions it might create :)
>> >>
>> >
>> > This sounds like a new dt binding that says the assigned parent should
>> > never change. It sounds sort of like gpio hogs. A clock-hogs binding?
>>
>> Ideally we want the clock driver to be able to reparent clocks freely
>> to get the best rate. But we also need some control over that to stop
>> consumers from being reparented in undesired ways. Eg. you might want
>> to make sure the GPU gets its own PLL so it can be reclocked easily,
>> and putting another device on the GPU's PLL could prevent that.
>>
>> The only way to achieve this today is (1) never do any reparenting in
>> the clock driver; and (2) use assigned-clock-parents in the DT to set
>> up the entire clock tree manually.
>>
>> Maxime said that (2) is basically wrong -- if assigned-clock-parents
>> provides no guarantee on what the OS does "after boot" then the OS is
>> pretty much free to ignore it.
>
> I didn't really say it's wrong, just that it never provided the
> guarantee you expect it to provide. I can't really say whether it's an
> issue or not on your platform.
>
> It's mostly unrelated to this series though, none of these patches
> affect that behavior in one way or the other.
I know. Sorry for derailing your patch :(
>> My suggestion: add a per-clock bitmap to keep track of which parents
>> are allowed. Any operation that would select a parent clock not on the
>> whitelist should fail. Automatic reparenting should only select from
>> clocks on the whitelist. And we need new DT bindings for controlling
>> the whitelist, for example:
>>
>> clock-parents-0 = <&clk1>, <&pll_c>;
>> clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
>>
>> This means that clk1 can only have pll_c as a parent, while clk2 can
>> have pll_a or pll_b as parents. By default every clock will be able
>> to use any parent, so a list is only needed if the machine needs a
>> more restrictive policy.
>>
>> assigned-clock-parents should disable automatic reparenting, but allow
>> explicit clk_set_parent(). This will allow clock drivers to start doing
>> reparenting without breaking old DTs.
>
> I'm generally not a fan of putting all these policies in the device
> tree. Do you have an example where it wouldn't be possible to do exactly
> this from the driver itself?
>
> Maxime
I'm confused. What's implicit in the example is clk1 and clk2 might
have *other* possible choices of parent clock and the device tree is
limiting what the OS is allowed to choose.
Why would you put such arbitrary limitations into the driver? They
would be different from machine to machine, unless the clock tree is
so simple there is only *one* meaningful way to configure it. Most
SoCs are complicated enough that there will be tradeoffs depending
on what peripherals you are using (typically a single machine will
not use *every* peripheral device provided by the SoC).
On Fri, Mar 24, 2023 at 08:58:48PM +0000, Aidan MacDonald wrote:
> >> My suggestion: add a per-clock bitmap to keep track of which parents
> >> are allowed. Any operation that would select a parent clock not on the
> >> whitelist should fail. Automatic reparenting should only select from
> >> clocks on the whitelist. And we need new DT bindings for controlling
> >> the whitelist, for example:
> >>
> >> clock-parents-0 = <&clk1>, <&pll_c>;
> >> clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
> >>
> >> This means that clk1 can only have pll_c as a parent, while clk2 can
> >> have pll_a or pll_b as parents. By default every clock will be able
> >> to use any parent, so a list is only needed if the machine needs a
> >> more restrictive policy.
> >>
> >> assigned-clock-parents should disable automatic reparenting, but allow
> >> explicit clk_set_parent(). This will allow clock drivers to start doing
> >> reparenting without breaking old DTs.
> >
> > I'm generally not a fan of putting all these policies in the device
> > tree. Do you have an example where it wouldn't be possible to do exactly
> > this from the driver itself?
>
> I'm confused. What's implicit in the example is clk1 and clk2 might
> have *other* possible choices of parent clock and the device tree is
> limiting what the OS is allowed to choose.
>
> Why would you put such arbitrary limitations into the driver?
Why would we put such arbitrary limitations in the firmware? As this
entire thread can attest, people are already using the device tree to
work around the limitations of the Linux driver, or reduce the
features of Linux because they can rely on the device tree. Either
way, it's linked to the state of the Linux driver, and any other OS or
Linux version could very well implement something more dynamic.
> They would be different from machine to machine, unless the clock
> tree is so simple there is only *one* meaningful way to configure
> it.
If we look at the device trees we have in-tree, most of the users of
assigned-clocks are the same from one board to another.
> Most SoCs are complicated enough that there will be tradeoffs
> depending on what peripherals you are using (typically a single
> machine will not use *every* peripheral device provided by the SoC).
We already have APIs to lock parents or rates on a given clock from
the consumer. It's far superior (feature-wise) than what the device
tree will ever offer because it's code, and it depends on the usage
already since an unused driver won't probe.
Maxime
Hi Maxime,
Le lundi 27 mars 2023 à 21:24 +0200, Maxime Ripard a écrit :
> On Fri, Mar 24, 2023 at 08:58:48PM +0000, Aidan MacDonald wrote:
> > > > My suggestion: add a per-clock bitmap to keep track of which
> > > > parents
> > > > are allowed. Any operation that would select a parent clock not
> > > > on the
> > > > whitelist should fail. Automatic reparenting should only select
> > > > from
> > > > clocks on the whitelist. And we need new DT bindings for
> > > > controlling
> > > > the whitelist, for example:
> > > >
> > > > clock-parents-0 = <&clk1>, <&pll_c>;
> > > > clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
> > > >
> > > > This means that clk1 can only have pll_c as a parent, while
> > > > clk2 can
> > > > have pll_a or pll_b as parents. By default every clock will be
> > > > able
> > > > to use any parent, so a list is only needed if the machine
> > > > needs a
> > > > more restrictive policy.
> > > >
> > > > assigned-clock-parents should disable automatic reparenting,
> > > > but allow
> > > > explicit clk_set_parent(). This will allow clock drivers to
> > > > start doing
> > > > reparenting without breaking old DTs.
> > >
> > > I'm generally not a fan of putting all these policies in the
> > > device
> > > tree. Do you have an example where it wouldn't be possible to do
> > > exactly
> > > this from the driver itself?
> >
> > I'm confused. What's implicit in the example is clk1 and clk2 might
> > have *other* possible choices of parent clock and the device tree
> > is
> > limiting what the OS is allowed to choose.
> >
> > Why would you put such arbitrary limitations into the driver?
>
> Why would we put such arbitrary limitations in the firmware? As this
> entire thread can attest, people are already using the device tree to
> work around the limitations of the Linux driver, or reduce the
> features of Linux because they can rely on the device tree. Either
> way, it's linked to the state of the Linux driver, and any other OS
> or
> Linux version could very well implement something more dynamic.
Probably because if we have to choose between setting policy in the
kernel or in the firmware, it is arguably better to set it in the
firmware.
Especially when talking about clocks, as the firmware is already the
one programming the boot clocks.
Cheers,
-Paul
> > They would be different from machine to machine, unless the clock
> > tree is so simple there is only *one* meaningful way to configure
> > it.
>
> If we look at the device trees we have in-tree, most of the users of
> assigned-clocks are the same from one board to another.
>
> > Most SoCs are complicated enough that there will be tradeoffs
> > depending on what peripherals you are using (typically a single
> > machine will not use *every* peripheral device provided by the
> > SoC).
>
> We already have APIs to lock parents or rates on a given clock from
> the consumer. It's far superior (feature-wise) than what the device
> tree will ever offer because it's code, and it depends on the usage
> already since an unused driver won't probe.
>
> Maxime
On Wed, Apr 05, 2023 at 02:57:26PM +0200, Paul Cercueil wrote:
> Le lundi 27 mars 2023 ? 21:24 +0200, Maxime Ripard a ?crit?:
> > On Fri, Mar 24, 2023 at 08:58:48PM +0000, Aidan MacDonald wrote:
> > > > > My suggestion: add a per-clock bitmap to keep track of which
> > > > > parents
> > > > > are allowed. Any operation that would select a parent clock not
> > > > > on the
> > > > > whitelist should fail. Automatic reparenting should only select
> > > > > from
> > > > > clocks on the whitelist. And we need new DT bindings for
> > > > > controlling
> > > > > the whitelist, for example:
> > > > >
> > > > > ??? clock-parents-0 = <&clk1>, <&pll_c>;
> > > > > ??? clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
> > > > >
> > > > > This means that clk1 can only have pll_c as a parent, while
> > > > > clk2 can
> > > > > have pll_a or pll_b as parents. By default every clock will be
> > > > > able
> > > > > to use any parent, so a list is only needed if the machine
> > > > > needs a
> > > > > more restrictive policy.
> > > > >
> > > > > assigned-clock-parents should disable automatic reparenting,
> > > > > but allow
> > > > > explicit clk_set_parent(). This will allow clock drivers to
> > > > > start doing
> > > > > reparenting without breaking old DTs.
> > > >
> > > > I'm generally not a fan of putting all these policies in the
> > > > device
> > > > tree. Do you have an example where it wouldn't be possible to do
> > > > exactly
> > > > this from the driver itself?
> > >
> > > I'm confused. What's implicit in the example is clk1 and clk2 might
> > > have *other* possible choices of parent clock and the device tree
> > > is
> > > limiting what the OS is allowed to choose.
> > >
> > > Why would you put such arbitrary limitations into the driver?
> >
> > Why would we put such arbitrary limitations in the firmware? As this
> > entire thread can attest, people are already using the device tree to
> > work around the limitations of the Linux driver, or reduce the
> > features of Linux because they can rely on the device tree. Either
> > way, it's linked to the state of the Linux driver, and any other OS
> > or
> > Linux version could very well implement something more dynamic.
>
> Probably because if we have to choose between setting policy in the
> kernel or in the firmware, it is arguably better to set it in the
> firmware.
I have a very different view on this I guess. Firmware is (most of the
time) hard to update, and the policy depend on the state of support of a
given OS so it's likely to evolve. The kernel is the best place to me to
put that kind of policy. Why do you think differently?
> Especially when talking about clocks, as the firmware is already the
> one programming the boot clocks.
I'm not sure what your point is there. I don't think I ever saw a
firmware getting the clocks right for every possible scenario on a given
platform. And if it was indeed the case, then we wouldn't even a kernel
driver.
Maxime
Le mercredi 05 avril 2023 à 16:50 +0200, Maxime Ripard a écrit :
> On Wed, Apr 05, 2023 at 02:57:26PM +0200, Paul Cercueil wrote:
> > Le lundi 27 mars 2023 à 21:24 +0200, Maxime Ripard a écrit :
> > > On Fri, Mar 24, 2023 at 08:58:48PM +0000, Aidan MacDonald wrote:
> > > > > > My suggestion: add a per-clock bitmap to keep track of
> > > > > > which
> > > > > > parents
> > > > > > are allowed. Any operation that would select a parent clock
> > > > > > not
> > > > > > on the
> > > > > > whitelist should fail. Automatic reparenting should only
> > > > > > select
> > > > > > from
> > > > > > clocks on the whitelist. And we need new DT bindings for
> > > > > > controlling
> > > > > > the whitelist, for example:
> > > > > >
> > > > > > clock-parents-0 = <&clk1>, <&pll_c>;
> > > > > > clock-parents-1 = <&clk2>, <&pll_a>, <&pll_b>;
> > > > > >
> > > > > > This means that clk1 can only have pll_c as a parent, while
> > > > > > clk2 can
> > > > > > have pll_a or pll_b as parents. By default every clock will
> > > > > > be
> > > > > > able
> > > > > > to use any parent, so a list is only needed if the machine
> > > > > > needs a
> > > > > > more restrictive policy.
> > > > > >
> > > > > > assigned-clock-parents should disable automatic
> > > > > > reparenting,
> > > > > > but allow
> > > > > > explicit clk_set_parent(). This will allow clock drivers to
> > > > > > start doing
> > > > > > reparenting without breaking old DTs.
> > > > >
> > > > > I'm generally not a fan of putting all these policies in the
> > > > > device
> > > > > tree. Do you have an example where it wouldn't be possible to
> > > > > do
> > > > > exactly
> > > > > this from the driver itself?
> > > >
> > > > I'm confused. What's implicit in the example is clk1 and clk2
> > > > might
> > > > have *other* possible choices of parent clock and the device
> > > > tree
> > > > is
> > > > limiting what the OS is allowed to choose.
> > > >
> > > > Why would you put such arbitrary limitations into the driver?
> > >
> > > Why would we put such arbitrary limitations in the firmware? As
> > > this
> > > entire thread can attest, people are already using the device
> > > tree to
> > > work around the limitations of the Linux driver, or reduce the
> > > features of Linux because they can rely on the device tree.
> > > Either
> > > way, it's linked to the state of the Linux driver, and any other
> > > OS
> > > or
> > > Linux version could very well implement something more dynamic.
> >
> > Probably because if we have to choose between setting policy in the
> > kernel or in the firmware, it is arguably better to set it in the
> > firmware.
>
> I have a very different view on this I guess. Firmware is (most of
> the
> time) hard to update, and the policy depend on the state of support
> of a
> given OS so it's likely to evolve. The kernel is the best place to me
> to
> put that kind of policy. Why do you think differently?
Because the clocks configuration can be board-specific. And you don't
really want board-specific stuff in the driver.
If we take the Ingenic JZ4770 SoC as example, on one board we parent
everything we can to the PLL1 clock and set it to 432 MHz (the least
common multiple). Then the PLL0 (which drives the DDR and CPU clocks)
can be updated to overclock/underclock the CPU and RAM.
So should that be hardcoded in the driver? Well, for a different board,
for which overclock is not really needed, it's better to parent
everything to PLL0 so that PLL1 can be shut down to save power. So what
policy should be hardcoded in the driver?
>
> > Especially when talking about clocks, as the firmware is already
> > the
> > one programming the boot clocks.
>
> I'm not sure what your point is there. I don't think I ever saw a
> firmware getting the clocks right for every possible scenario on a
> given
> platform. And if it was indeed the case, then we wouldn't even a
> kernel
> driver.
My point is that there is already policy in how the firmware sets up
the hardware; and most often than not, the kernel driver won't change
that (e.g. you're probably not going to touch the DDR clock). It's
better to have all policy in the firmware then, instead of having some
in the firmware, and some in the kernel driver.
Cheers,
-Paul