Re: [PATCH 4/4] clk: implement sync_state support
From: Sverdlin, Alexander
Date: Tue Sep 22 2026 - 09:36:51 EST
Hi Brian, Alexandre,
thanks for looking into this! Alexandre asked me to confirm that
it will work for RTC devices, but my current understanding is that
it will not and here is why:
On Fri, 2026-06-26 at 12:32 -0400, Brian Masney wrote:
> The existing support for disabling unused clks runs in the late initcall
> stage, and it has been known for a long time that this is broken since
> it runs too early in the boot up process. It doesn't work for kernel
> modules, and it also doesn't work if all of the consumers haven't fully
> probed yet. Folks have long recommended to boot certain platforms with
> clk_ignore_unused to work around issues with disabling unused clks.
>
> Let's go ahead and add a framework-level sync_state callback for the clk
> subsystem. If a driver doesn't have a sync_state callback configured,
> which is the 99+% use case today, then let's set it up to use the
> clk_sync_state() introduced in this commit so that no driver changes
> are needed.
>
> At the time of this writing, there are currently only 7 clk drivers that
> implement sync_state, and all are Qualcomm SoCs where they interact with
> the interconnect framework via icc_sync_state(). A shared helper has
> been created for this platform that calls clk_sync_state(). It is
> expected that any new clk drivers that want to implement their own
> sync_state will also need to call clk_sync_state() at the end of their
> custom sync_state callback.
>
> There will be several stages of disabling unused clks:
>
> - The first phase will be executed at late_initcall and it will only
> disable unused clks that do not have a struct dev.
> - The sync_state callback will be invoked for each clk driver once all
> consumers have probed.
>
> This is based on previous attempts by Saravana Kannan and Abel Vesa
> that are linked below.
>
> This change was tested on a Thinkpad x13s laptop.
>
> [ 0.366737] clk: Disabling unused clocks not associated with a device
> [ 0.367232] PM: genpd: Disabling unused power domains
> [ 7.791413] qcom-qmp-pcie-phy 1c24000.phy: clk: Disabling unused clocks
> [ 7.799702] qcom_aoss_qmp c300000.power-management: clk: Disabling unused clocks
> [ 8.548820] qcom-qmp-pcie-phy 1c14000.phy: clk: Disabling unused clocks
> [ 9.121849] qcom-qmp-usb-phy 88f1000.phy: clk: Disabling unused clocks
> [ 9.121985] qcom-qmp-usb-phy 88ef000.phy: clk: Disabling unused clocks
> [ 9.122691] qcom-edp-phy aec5a00.phy: clk: Disabling unused clocks
> [ 9.122760] disp_cc-sc8280xp af00000.clock-controller: clk: Disabling unused clocks
> [ 9.142121] qcom-qmp-combo-phy 88eb000.phy: clk: Disabling unused clocks
> [ 9.169149] qcom-qmp-combo-phy 8903000.phy: clk: Disabling unused clocks
> [ 16.057997] qcom-cpufreq-hw 18591000.cpufreq: clk: Disabling unused clocks
> [ 16.058149] clk-rpmh 18200000.rsc:clock-controller: clk: Disabling unused clocks
> [ 16.334879] qcom-qmp-pcie-phy 1c06000.phy: clk: Disabling unused clocks
> [ 16.706113] camcc-sc8280xp ad00000.clock-controller: clk: Disabling unused clocks
> [ 21.565731] q6prm-lpass-clock 3000000.remoteproc:glink-edge:gpr:service@2:clock-controller: clk: Disabling unused clocks
> [ 21.597069] va_macro 3370000.codec: clk: Disabling unused clocks
> [ 21.605039] rx_macro 3200000.rxmacro: clk: Disabling unused clocks
> [ 21.630313] wsa_macro 3240000.codec: clk: Disabling unused clocks
> [ 21.635069] tx_macro 3220000.txmacro: clk: Disabling unused clocks
>
> Tested-by: Jens Glathe <jens.glathe@xxxxxxxxxxxxxxxxxxxxxx>
> Tested-by: Neil Armstrong <neil.armstrong@xxxxxxxxxx> # on AML-S905X-CC, AML-S805X-CC, BPI-M2S, BPI-M5
> Tested-by: Xuyang Dong <dongxuyang@xxxxxxxxxxxxxxxxxx> # hfp550
> Signed-off-by: Brian Masney <bmasney@xxxxxxxxxx>
> Link: https://www.youtube.com/watch?v=tXYzM8yLIQA
> Link: https://lore.kernel.org/all/20210407034456.516204-1-saravanak@xxxxxxxxxx/
> Link: https://lore.kernel.org/all/20221227204528.1899863-1-abel.vesa@xxxxxxxxxx/
> ---
> drivers/clk/clk.c | 78 ++++++++++++++++++++++++++++++++++++++++++++---------
> include/linux/clk.h | 14 ++++++++++
> 2 files changed, 79 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index 048adfa86a5d..2b6fae54702a 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -1442,14 +1442,39 @@ static void clk_core_disable_unprepare(struct clk_core *core)
> clk_core_unprepare_lock(core);
> }
>
> -static void __init clk_unprepare_unused_subtree(struct clk_core *core)
> +/*
> + * Returns true if @core should be skipped during an unused-clock sweep for
> + * @dev. When @dev is NULL the sweep is the global late_initcall pass; when
> + * @dev is non-NULL the sweep is a per-device sync_state pass.
> + */
> +static bool clk_core_skip_unused(struct clk_core *core, struct device *dev)
> +{
> + /*
> + * At late_initcall, skip clocks that belong to a device — they will be
> + * handled at sync_state time.
> + */
> + if (!dev && core->dev)
> + return true;
This will skip all core->dev != NULL, but the deferred sync-based disabling
will only work for those driver-bound devices, not driverless class devices?
Probably a regression in power consumption as-is, see also below...
> +
> + /* When called from sync_state, only process clocks for this device. */
> + if (dev && core->dev != dev)
> + return true;
> +
> + return false;
> +}
> +
> +static void clk_unprepare_unused_subtree(struct clk_core *core,
> + struct device *dev)
> {
> struct clk_core *child;
>
> lockdep_assert_held(&prepare_lock);
>
> hlist_for_each_entry(child, &core->children, child_node)
> - clk_unprepare_unused_subtree(child);
> + clk_unprepare_unused_subtree(child, dev);
> +
> + if (clk_core_skip_unused(core, dev))
> + return;
>
> if (core->prepare_count)
> return;
> @@ -1467,7 +1492,8 @@ static void __init clk_unprepare_unused_subtree(struct clk_core *core)
> }
> }
>
> -static void __init clk_disable_unused_subtree(struct clk_core *core)
> +static void clk_disable_unused_subtree(struct clk_core *core,
> + struct device *dev)
> {
> struct clk_core *child;
> unsigned long flags;
> @@ -1475,7 +1501,10 @@ static void __init clk_disable_unused_subtree(struct clk_core *core)
> lockdep_assert_held(&prepare_lock);
>
> hlist_for_each_entry(child, &core->children, child_node)
> - clk_disable_unused_subtree(child);
> + clk_disable_unused_subtree(child, dev);
> +
> + if (clk_core_skip_unused(core, dev))
> + return;
>
> if (core->flags & CLK_OPS_PARENT_ENABLE)
> clk_core_prepare_enable(core->parent);
> @@ -1508,7 +1537,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core)
> clk_core_disable_unprepare(core->parent);
> }
>
> -static bool clk_ignore_unused __initdata;
> +static bool clk_ignore_unused;
> static int __init clk_ignore_unused_setup(char *__unused)
> {
> clk_ignore_unused = true;
> @@ -1516,17 +1545,27 @@ static int __init clk_ignore_unused_setup(char *__unused)
> }
> __setup("clk_ignore_unused", clk_ignore_unused_setup);
>
> -static int __init clk_disable_unused(void)
> +static int __clk_disable_unused(struct device *dev)
> {
> struct clk_core *core;
> int ret;
>
> if (clk_ignore_unused) {
> - pr_warn("clk: Not disabling unused clocks\n");
> + /*
> + * Only show this message once for the providers that are not
> + * associated with a device, which is the first pass of
> + * disabling unused clocks. Otherwise it'll show up for each
> + * clk provider as well.
> + */
> + if (!dev)
> + pr_warn("clk: Not disabling unused clocks\n");
> return 0;
> }
>
> - pr_info("clk: Disabling unused clocks\n");
> + if (dev)
> + dev_info(dev, "clk: Disabling unused clocks\n");
It becomes quite chatty on large systems?
> + else
> + pr_info("clk: Disabling unused clocks not associated with a device\n");
>
> ret = clk_pm_runtime_get_all();
> if (ret)
> @@ -1538,16 +1577,16 @@ static int __init clk_disable_unused(void)
> clk_prepare_lock();
>
> hlist_for_each_entry(core, &clk_root_list, child_node)
> - clk_disable_unused_subtree(core);
> + clk_disable_unused_subtree(core, dev);
>
> hlist_for_each_entry(core, &clk_orphan_list, child_node)
> - clk_disable_unused_subtree(core);
> + clk_disable_unused_subtree(core, dev);
>
> hlist_for_each_entry(core, &clk_root_list, child_node)
> - clk_unprepare_unused_subtree(core);
> + clk_unprepare_unused_subtree(core, dev);
>
> hlist_for_each_entry(core, &clk_orphan_list, child_node)
> - clk_unprepare_unused_subtree(core);
> + clk_unprepare_unused_subtree(core, dev);
>
> clk_prepare_unlock();
>
> @@ -1555,8 +1594,19 @@ static int __init clk_disable_unused(void)
>
> return 0;
> }
> +
> +static int __init clk_disable_unused(void)
> +{
> + return __clk_disable_unused(NULL);
> +}
> late_initcall_sync(clk_disable_unused);
>
> +void clk_sync_state(struct device *dev)
> +{
> + __clk_disable_unused(dev);
> +}
> +EXPORT_SYMBOL_GPL(clk_sync_state);
> +
> static int clk_core_determine_round_nolock(struct clk_core *core,
> struct clk_rate_request *req)
> {
> @@ -4339,8 +4389,10 @@ __clk_register(struct device *dev, struct device_node *np, struct clk_hw *hw)
> core->dev = dev;
> clk_pm_runtime_init(core);
> core->of_node = np;
> - if (dev && dev->driver)
> + if (dev && dev->driver) {
So this will only work for driver-bound devices, but not for RTC-class
devices, right?
In the use-case we've discussed with Alexandre:
clk = devm_clk_register(&pcf85063->rtc->dev, &pcf85063->clkout_hw);
&pcf85063->rtc->dev is an rtc_class device -- it never has dev->driver set.
> core->owner = dev->driver->owner;
> + dev_add_sync_state(dev, clk_sync_state);
> + }
> core->hw = hw;
> core->flags = init->flags;
> core->num_parents = init->num_parents;
> diff --git a/include/linux/clk.h b/include/linux/clk.h
> index 998ba3f261da..31a0c9224c46 100644
> --- a/include/linux/clk.h
> +++ b/include/linux/clk.h
> @@ -846,6 +846,20 @@ void devm_clk_put(struct device *dev, struct clk *clk);
> */
>
>
> +/*
> + * clk_sync_state - sync_state callback to disable unused clocks
> + * @dev: the clock provider device whose unused clocks should be disabled
> + *
> + * It is called by the driver core once all consumers of @dev have probed,
> + * and disables any clocks belonging to @dev that are unused at that point.
> + *
> + * If a clock provider doesn't have a sync_state callback, then the framework
> + * will set up clk_sync_state() on your drivers behalf. If your driver needs
That is only true if dev->driver is set at __clk_register() time?
> + * a sync_state callback, then that callback also needs to call
> + * clk_sync_state().
> + */
> +void clk_sync_state(struct device *dev);
> +
> /**
> * clk_round_rate - adjust a rate to the exact rate a clock can provide
> * @clk: clock source
I've also tested the series with regards to PCF85053 RTC clkout issue, but
the clock is not being disabled in practice.
--
Alexander Sverdlin
Siemens AG
www.siemens.com