Message ID | 20241004133953.494445-1-jbrunet@baylibre.com (mailing list archive) |
---|---|
State | New, archived |
Delegated to: | Neil Armstrong |
Headers | show |
Series | [RFC] clk: core: refine disable unused clocks | expand |
hi Jerome: Tranks for your REF. I looked at your patch and there are some parts that I don't quite understand: The original intention of CLK_OPS_PARENT_ENABLE was to solve the issue of "parents need enable _during _gate/ungate, set rate and re-parent" when setting a clock. After setting the clock, it can still be disabled. However, from what I see in your patch, the handling logic seems more like "parents need _always _ gate during clock gate period"? On 10/4/2024 9:39 PM, Jerome Brunet wrote: > [ EXTERNAL EMAIL ] > > As it as been pointed out numerous times, flagging a clock with > CLK_IGNORE_UNUSED does _not_ guarantee that clock left enabled will stay > on. The clock will get disabled if any enable/disable cycle happens on it > or its parent clocks. > > Because enable/disable cycles will disable unused clocks, > clk_disable_unused() should not trigger such cycle. Doing so disregard > the flag if set for any parent clocks. This is problematic with > CLK_OPS_PARENT_ENABLE handling. > > To solve this, and a couple other issues, pass the parent status to the > child while walking the subtree, and return whether child ignored disable, > or not. > > * Knowing the parent status allows to safely disable clocks with > CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means > that, while the clock is not gated it is effectively disabled. Turning on > the parents to sanitize the sitation would bring back our initial > problem, so just let it sanitize itself when the clock gets used. > > * If a clock is not actively used (enabled_count == 0), does not have > CLK_IGNORE_UNUSED but the hw enabled all the way to the root clock, and a > child ignored the disable, it should ignore the disable too. Doing so > avoids disabling what is feading the children. Let the flag trickle down > the tree. This has the added benefit to transfer the information to the > unprepare path, so we don't unprepare the parent of a clock that ignored > a disable. > > * An enabled clock must be prepared in CCF but we can't rely solely on > counts at clk_disable_unused() stage. Make sure an enabled clock is > considered prepared too, even if does not implement the related callback. > Also make sure only disabled clocks get unprepared. > > Signed-off-by: Jerome Brunet <jbrunet@baylibre.com> > --- > > This is sent as an RFC to continue the discussion started by Chuan. > It is not meant to be applied as it is. > > > drivers/clk/clk.c | 92 ++++++++++++++++++++++++++++++----------------- > 1 file changed, 60 insertions(+), 32 deletions(-) > > diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c > index d02451f951cf..41c4504a41f1 100644 > --- a/drivers/clk/clk.c > +++ b/drivers/clk/clk.c > @@ -332,17 +332,6 @@ static bool clk_core_is_enabled(struct clk_core *core) > } > } > > - /* > - * This could be called with the enable lock held, or from atomic > - * context. If the parent isn't enabled already, we can't do > - * anything here. We can also assume this clock isn't enabled. > - */ > - if ((core->flags & CLK_OPS_PARENT_ENABLE) && core->parent) This judgment of CLK_OPS_PARENT_ENABLE seems redundant. According to normal logic, if the parent is disabled, its children will also be forced to disable. This seems unrelated to whether CLK_OPS_PARENT_ENABLE is configured.
On Fri 08 Nov 2024 at 15:59, Chuan Liu <chuan.liu@amlogic.com> wrote: > hi Jerome: > > Tranks for your REF. I looked at your patch and there are some parts > that I don't quite understand: The original intention of > CLK_OPS_PARENT_ENABLE was to solve the issue of "parents need enable > _during _gate/ungate, set rate and re-parent" when setting a clock. After > setting the clock, it can still be disabled. However, from what I see in > your patch, the handling logic seems more like "parents need _always _ gate > during clock gate period"? As explained in the description, the problem with CLK_IGNORE_UNUSED and CLK_OPS_PARENT_ENABLE is that you'll get cycle of enable/disable, which will disable any parent clock that may have a been enabled and expected to be ignored. IOW, the CCF changes the state of the tree while inspecting it. This change solves that. > > On 10/4/2024 9:39 PM, Jerome Brunet wrote: >> [ EXTERNAL EMAIL ] >> >> As it as been pointed out numerous times, flagging a clock with >> CLK_IGNORE_UNUSED does _not_ guarantee that clock left enabled will stay >> on. The clock will get disabled if any enable/disable cycle happens on it >> or its parent clocks. >> >> Because enable/disable cycles will disable unused clocks, >> clk_disable_unused() should not trigger such cycle. Doing so disregard >> the flag if set for any parent clocks. This is problematic with >> CLK_OPS_PARENT_ENABLE handling. >> >> To solve this, and a couple other issues, pass the parent status to the >> child while walking the subtree, and return whether child ignored disable, >> or not. >> >> * Knowing the parent status allows to safely disable clocks with >> CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means >> that, while the clock is not gated it is effectively disabled. Turning on >> the parents to sanitize the sitation would bring back our initial >> problem, so just let it sanitize itself when the clock gets used. >> >> * If a clock is not actively used (enabled_count == 0), does not have >> CLK_IGNORE_UNUSED but the hw enabled all the way to the root clock, and a >> child ignored the disable, it should ignore the disable too. Doing so >> avoids disabling what is feading the children. Let the flag trickle down >> the tree. This has the added benefit to transfer the information to the >> unprepare path, so we don't unprepare the parent of a clock that ignored >> a disable. >> >> * An enabled clock must be prepared in CCF but we can't rely solely on >> counts at clk_disable_unused() stage. Make sure an enabled clock is >> considered prepared too, even if does not implement the related callback. >> Also make sure only disabled clocks get unprepared. >> >> Signed-off-by: Jerome Brunet <jbrunet@baylibre.com> >> --- >> >> This is sent as an RFC to continue the discussion started by Chuan. >> It is not meant to be applied as it is. >> >> >> drivers/clk/clk.c | 92 ++++++++++++++++++++++++++++++----------------- >> 1 file changed, 60 insertions(+), 32 deletions(-) >> >> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c >> index d02451f951cf..41c4504a41f1 100644 >> --- a/drivers/clk/clk.c >> +++ b/drivers/clk/clk.c >> @@ -332,17 +332,6 @@ static bool clk_core_is_enabled(struct clk_core *core) >> } >> } >> >> - /* >> - * This could be called with the enable lock held, or from atomic >> - * context. If the parent isn't enabled already, we can't do >> - * anything here. We can also assume this clock isn't enabled. >> - */ >> - if ((core->flags & CLK_OPS_PARENT_ENABLE) && core->parent) > > This judgment of CLK_OPS_PARENT_ENABLE seems redundant. According to > normal logic, if the parent is disabled, its children will also be > forced to disable. This seems unrelated to whether CLK_OPS_PARENT_ENABLE > is configured.
On 11/8/2024 4:38 PM, Jerome Brunet wrote: > [ EXTERNAL EMAIL ] > > On Fri 08 Nov 2024 at 15:59, Chuan Liu <chuan.liu@amlogic.com> wrote: > >> hi Jerome: >> >> Tranks for your REF. I looked at your patch and there are some parts >> that I don't quite understand: The original intention of >> CLK_OPS_PARENT_ENABLE was to solve the issue of "parents need enable >> _during _gate/ungate, set rate and re-parent" when setting a clock. After >> setting the clock, it can still be disabled. However, from what I see in >> your patch, the handling logic seems more like "parents need _always _ gate >> during clock gate period"? > As explained in the description, the problem with CLK_IGNORE_UNUSED and > CLK_OPS_PARENT_ENABLE is that you'll get cycle of enable/disable, which > will disable any parent clock that may have a been enabled and expected > to be ignored. > > IOW, the CCF changes the state of the tree while inspecting it. > This change solves that. Ok, I understand your idea now... I have gotten myself tangled up. >> On 10/4/2024 9:39 PM, Jerome Brunet wrote: >>> [ EXTERNAL EMAIL ] >>> >>> As it as been pointed out numerous times, flagging a clock with >>> CLK_IGNORE_UNUSED does _not_ guarantee that clock left enabled will stay >>> on. The clock will get disabled if any enable/disable cycle happens on it >>> or its parent clocks. >>> >>> Because enable/disable cycles will disable unused clocks, >>> clk_disable_unused() should not trigger such cycle. Doing so disregard >>> the flag if set for any parent clocks. This is problematic with >>> CLK_OPS_PARENT_ENABLE handling. >>> >>> To solve this, and a couple other issues, pass the parent status to the >>> child while walking the subtree, and return whether child ignored disable, >>> or not. >>> >>> * Knowing the parent status allows to safely disable clocks with >>> CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means >>> that, while the clock is not gated it is effectively disabled. Turning on >>> the parents to sanitize the sitation would bring back our initial >>> problem, so just let it sanitize itself when the clock gets used. >>> >>> * If a clock is not actively used (enabled_count == 0), does not have >>> CLK_IGNORE_UNUSED but the hw enabled all the way to the root clock, and a >>> child ignored the disable, it should ignore the disable too. Doing so >>> avoids disabling what is feading the children. Let the flag trickle down >>> the tree. This has the added benefit to transfer the information to the >>> unprepare path, so we don't unprepare the parent of a clock that ignored >>> a disable. >>> >>> * An enabled clock must be prepared in CCF but we can't rely solely on >>> counts at clk_disable_unused() stage. Make sure an enabled clock is >>> considered prepared too, even if does not implement the related callback. >>> Also make sure only disabled clocks get unprepared. >>> >>> Signed-off-by: Jerome Brunet <jbrunet@baylibre.com> >>> --- >>> >>> This is sent as an RFC to continue the discussion started by Chuan. >>> It is not meant to be applied as it is. >>> >>> >>> drivers/clk/clk.c | 92 ++++++++++++++++++++++++++++++----------------- >>> 1 file changed, 60 insertions(+), 32 deletions(-) >>> >>> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c >>> index d02451f951cf..41c4504a41f1 100644 >>> --- a/drivers/clk/clk.c >>> +++ b/drivers/clk/clk.c >>> @@ -332,17 +332,6 @@ static bool clk_core_is_enabled(struct clk_core *core) >>> } >>> } >>> >>> - /* >>> - * This could be called with the enable lock held, or from atomic >>> - * context. If the parent isn't enabled already, we can't do >>> - * anything here. We can also assume this clock isn't enabled. >>> - */ >>> - if ((core->flags & CLK_OPS_PARENT_ENABLE) && core->parent) >> This judgment of CLK_OPS_PARENT_ENABLE seems redundant. According to >> normal logic, if the parent is disabled, its children will also be >> forced to disable. This seems unrelated to whether CLK_OPS_PARENT_ENABLE >> is configured.
On Fri 08 Nov 2024 at 17:23, Chuan Liu <chuan.liu@amlogic.com> wrote: >>>> - if (core->flags & CLK_IGNORE_UNUSED) >>>> + /* >>>> + * If the parent is disabled but the gate is open, we should sanitize >>>> + * the situation. This will avoid an unexpected enable of the clock as >>>> + * soon as the parent is enabled, without control of CCF. >>>> + * >>>> + * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without >>>> + * forcefully enabling a whole part of the subtree. Just let the >>>> + * situation resolve it self on the first enable of the clock >>>> + */ >>>> + if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE)) > > At first, I couldn't grasp the logic behind the 'return' here. Now it's > clear. This approach is equivalent to completely giving up on > handling clocks with CLK_OPS_PARENT_ENABLE feature in > clk_disable_unused_subtree(). > No. It's handled correctly as long as the tree is in coherent state. What is not done anymore is fixing up an inconsistent tree, by this I mean: A clock with CLK_OPS_PARENT_ENABLE, which report enabled from its own registers but has its parent disabled. In that particular case, clk_disable_unused_subtree() won't be turning on everything to properly disable that one clock. That is the root cause of the problem you reported initially. The clock is disabled anyway. Every other case are properly handled (at least I think). >>>> goto unlock_out; >>>> >>>> /* >>>> @@ -1516,8 +1545,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core) >>>> >>>> unlock_out: >>>> clk_enable_unlock(flags); >>>> - if (core->flags & CLK_OPS_PARENT_ENABLE) >>>> - clk_core_disable_unprepare(core->parent); >>>> + return (core->flags & CLK_IGNORE_UNUSED) && enabled; >>>> } >>>> >>>> static bool clk_ignore_unused __initdata; >>>> @@ -1550,16 +1578,16 @@ 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, true); >>>> >>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>> - clk_disable_unused_subtree(core); >>>> + clk_disable_unused_subtree(core, true); >>>> >>>> hlist_for_each_entry(core, &clk_root_list, child_node) >>>> - clk_unprepare_unused_subtree(core); >>>> + clk_unprepare_unused_subtree(core, true); >>>> >>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>> - clk_unprepare_unused_subtree(core); >>>> + clk_unprepare_unused_subtree(core, true); >>>> >>>> clk_prepare_unlock(); >>>> >>>> -- >>>> 2.45.2 >>>> >> -- >> Jerome
On 11/8/2024 5:59 PM, Jerome Brunet wrote: > [ EXTERNAL EMAIL ] > > On Fri 08 Nov 2024 at 17:23, Chuan Liu <chuan.liu@amlogic.com> wrote: > >>>>> - if (core->flags & CLK_IGNORE_UNUSED) >>>>> + /* >>>>> + * If the parent is disabled but the gate is open, we should sanitize >>>>> + * the situation. This will avoid an unexpected enable of the clock as >>>>> + * soon as the parent is enabled, without control of CCF. >>>>> + * >>>>> + * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without >>>>> + * forcefully enabling a whole part of the subtree. Just let the >>>>> + * situation resolve it self on the first enable of the clock >>>>> + */ >>>>> + if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE)) >> At first, I couldn't grasp the logic behind the 'return' here. Now it's >> clear. This approach is equivalent to completely giving up on >> handling clocks with CLK_OPS_PARENT_ENABLE feature in >> clk_disable_unused_subtree(). >> > No. It's handled correctly as long as the tree is in coherent state. > > What is not done anymore is fixing up an inconsistent tree, by this I > mean: A clock with CLK_OPS_PARENT_ENABLE, which report enabled from its > own registers but has its parent disabled. > > In that particular case, clk_disable_unused_subtree() won't be turning on > everything to properly disable that one clock. That is the root cause of > the problem you reported initially. The clock is disabled anyway. > > Every other case are properly handled (at least I think). name en_sts flags clk_a 1 CLK_IGNORE_UNUSED clk_b 0 0 clk_c 1 CLK_OPS_PARENT_ENABLE Based on the above case: 1. When 'clk_c' is configured with CLK_OPS_PARENT_ENABLE, disabling 'clk_c' requires enabling 'clk_b' first (disabling 'clk_c' before disabling 'clk_b'). How can to ensure that during the period of disabling 'clk_c', 'clk_b' remains enabled? 2. 'clk_c' is not configured with CLK_IGNORE_UNUSED, it should be disabled later. However, here it goes to a 'goto' statement and then return 'false', ultimately resulting in 'clk_c' not being disabled? >>>>> goto unlock_out; >>>>> >>>>> /* >>>>> @@ -1516,8 +1545,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core) >>>>> >>>>> unlock_out: >>>>> clk_enable_unlock(flags); >>>>> - if (core->flags & CLK_OPS_PARENT_ENABLE) >>>>> - clk_core_disable_unprepare(core->parent); >>>>> + return (core->flags & CLK_IGNORE_UNUSED) && enabled; >>>>> } >>>>> >>>>> static bool clk_ignore_unused __initdata; >>>>> @@ -1550,16 +1578,16 @@ 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, true); >>>>> >>>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>>> - clk_disable_unused_subtree(core); >>>>> + clk_disable_unused_subtree(core, true); >>>>> >>>>> hlist_for_each_entry(core, &clk_root_list, child_node) >>>>> - clk_unprepare_unused_subtree(core); >>>>> + clk_unprepare_unused_subtree(core, true); >>>>> >>>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>>> - clk_unprepare_unused_subtree(core); >>>>> + clk_unprepare_unused_subtree(core, true); >>>>> >>>>> clk_prepare_unlock(); >>>>> >>>>> -- >>>>> 2.45.2 >>>>> >>> -- >>> Jerome > -- > Jerome
On Fri 08 Nov 2024 at 19:49, Chuan Liu <chuan.liu@amlogic.com> wrote: > On 11/8/2024 5:59 PM, Jerome Brunet wrote: >> [ EXTERNAL EMAIL ] >> >> On Fri 08 Nov 2024 at 17:23, Chuan Liu <chuan.liu@amlogic.com> wrote: >> >>>>>> - if (core->flags & CLK_IGNORE_UNUSED) >>>>>> + /* >>>>>> + * If the parent is disabled but the gate is open, we should sanitize >>>>>> + * the situation. This will avoid an unexpected enable of the clock as >>>>>> + * soon as the parent is enabled, without control of CCF. >>>>>> + * >>>>>> + * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without >>>>>> + * forcefully enabling a whole part of the subtree. Just let the >>>>>> + * situation resolve it self on the first enable of the clock >>>>>> + */ >>>>>> + if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE)) >>> At first, I couldn't grasp the logic behind the 'return' here. Now it's >>> clear. This approach is equivalent to completely giving up on >>> handling clocks with CLK_OPS_PARENT_ENABLE feature in >>> clk_disable_unused_subtree(). >>> >> No. It's handled correctly as long as the tree is in coherent state. >> >> What is not done anymore is fixing up an inconsistent tree, by this I >> mean: A clock with CLK_OPS_PARENT_ENABLE, which report enabled from its >> own registers but has its parent disabled. >> >> In that particular case, clk_disable_unused_subtree() won't be turning on >> everything to properly disable that one clock. That is the root cause of >> the problem you reported initially. The clock is disabled anyway. >> >> Every other case are properly handled (at least I think). > > name en_sts flags > clk_a 1 CLK_IGNORE_UNUSED > clk_b 0 0 > clk_c 1 CLK_OPS_PARENT_ENABLE > > Based on the above case: > 1. When 'clk_c' is configured with CLK_OPS_PARENT_ENABLE, disabling > 'clk_c' requires enabling 'clk_b' first (disabling 'clk_c' before > disabling 'clk_b'). How can to ensure that during the period of > disabling 'clk_c', 'clk_b' remains enabled? That's perfect example of incoherent state. How can 'clk_c' be enabled if its parent is disable. That makes no sense, so there is no point enabling a whole subtree for this IMO. > > 2. 'clk_c' is not configured with CLK_IGNORE_UNUSED, it should be > disabled later. However, here it goes to a 'goto' statement and then > return 'false', ultimately resulting in 'clk_c' not being disabled? We've discussed that 2 times already. This discussion is going in circles now. > >>>>>> goto unlock_out; >>>>>> >>>>>> /* >>>>>> @@ -1516,8 +1545,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core) >>>>>> >>>>>> unlock_out: >>>>>> clk_enable_unlock(flags); >>>>>> - if (core->flags & CLK_OPS_PARENT_ENABLE) >>>>>> - clk_core_disable_unprepare(core->parent); >>>>>> + return (core->flags & CLK_IGNORE_UNUSED) && enabled; >>>>>> } >>>>>> >>>>>> static bool clk_ignore_unused __initdata; >>>>>> @@ -1550,16 +1578,16 @@ 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, true); >>>>>> >>>>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>>>> - clk_disable_unused_subtree(core); >>>>>> + clk_disable_unused_subtree(core, true); >>>>>> >>>>>> hlist_for_each_entry(core, &clk_root_list, child_node) >>>>>> - clk_unprepare_unused_subtree(core); >>>>>> + clk_unprepare_unused_subtree(core, true); >>>>>> >>>>>> hlist_for_each_entry(core, &clk_orphan_list, child_node) >>>>>> - clk_unprepare_unused_subtree(core); >>>>>> + clk_unprepare_unused_subtree(core, true); >>>>>> >>>>>> clk_prepare_unlock(); >>>>>> >>>>>> -- >>>>>> 2.45.2 >>>>>> >>>> -- >>>> Jerome >> -- >> Jerome
On 11/12/2024 4:36 PM, Jerome Brunet wrote: > [ EXTERNAL EMAIL ] > > On Fri 08 Nov 2024 at 19:49, Chuan Liu <chuan.liu@amlogic.com> wrote: > >> On 11/8/2024 5:59 PM, Jerome Brunet wrote: >>> [ EXTERNAL EMAIL ] >>> >>> On Fri 08 Nov 2024 at 17:23, Chuan Liu <chuan.liu@amlogic.com> wrote: >>> >>>>>>> - if (core->flags & CLK_IGNORE_UNUSED) >>>>>>> + /* >>>>>>> + * If the parent is disabled but the gate is open, we should sanitize >>>>>>> + * the situation. This will avoid an unexpected enable of the clock as >>>>>>> + * soon as the parent is enabled, without control of CCF. >>>>>>> + * >>>>>>> + * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without >>>>>>> + * forcefully enabling a whole part of the subtree. Just let the >>>>>>> + * situation resolve it self on the first enable of the clock >>>>>>> + */ >>>>>>> + if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE)) >>>> At first, I couldn't grasp the logic behind the 'return' here. Now it's >>>> clear. This approach is equivalent to completely giving up on >>>> handling clocks with CLK_OPS_PARENT_ENABLE feature in >>>> clk_disable_unused_subtree(). Referring to the situation of 'clk_c' below, combined with your previous explanation: * Knowing the parent status allows to safely disable clocks with CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means that, while the clock is not gated it is effectively disabled. Turning on the parents to sanitize the sitation would bring back our initial problem, so just let it sanitize itself when the clock gets used. Do you mean 'clk_c' cases should be sanitized before clk_disable_unused() (such as during driver probe(), etc.)? Dropped in clk_disable_unused_subtree()? This is actually my biggest confusion.
diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c index d02451f951cf..41c4504a41f1 100644 --- a/drivers/clk/clk.c +++ b/drivers/clk/clk.c @@ -332,17 +332,6 @@ static bool clk_core_is_enabled(struct clk_core *core) } } - /* - * This could be called with the enable lock held, or from atomic - * context. If the parent isn't enabled already, we can't do - * anything here. We can also assume this clock isn't enabled. - */ - if ((core->flags & CLK_OPS_PARENT_ENABLE) && core->parent) - if (!clk_core_is_enabled(core->parent)) { - ret = false; - goto done; - } - ret = core->ops->is_enabled(core->hw); done: if (core->rpm_enabled) @@ -1454,22 +1443,39 @@ 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) +static bool __init clk_unprepare_unused_subtree(struct clk_core *core, + bool parent_prepared) { struct clk_core *child; + bool prepared; lockdep_assert_held(&prepare_lock); + /* + * Relying on count is not possible at this stage, so consider + * prepared an enabled clock, in case only .is_enabled() is + * implemented + */ + if (parent_prepared) + prepared = (clk_core_is_prepared(core) || + clk_core_is_enabled(core)); + else + prepared = false; + hlist_for_each_entry(child, &core->children, child_node) - clk_unprepare_unused_subtree(child); + if (clk_unprepare_unused_subtree(child, prepared) && + prepared && !core->prepare_count) + core->flags |= CLK_IGNORE_UNUSED; - if (core->prepare_count) - return; + if (core->flags & CLK_IGNORE_UNUSED || core->prepare_count) + goto out; - if (core->flags & CLK_IGNORE_UNUSED) - return; + if (!parent_prepared && (core->flags & CLK_OPS_PARENT_ENABLE)) + goto out; - if (clk_core_is_prepared(core)) { + /* Do not unprepare an enabled clock */ + if (clk_core_is_prepared(core) && + !clk_core_is_enabled(core)) { trace_clk_unprepare(core); if (core->ops->unprepare_unused) core->ops->unprepare_unused(core->hw); @@ -1477,27 +1483,50 @@ static void __init clk_unprepare_unused_subtree(struct clk_core *core) core->ops->unprepare(core->hw); trace_clk_unprepare_complete(core); } + +out: + return (core->flags & CLK_IGNORE_UNUSED) && prepared; } -static void __init clk_disable_unused_subtree(struct clk_core *core) +static bool __init clk_disable_unused_subtree(struct clk_core *core, + bool parent_enabled) { struct clk_core *child; unsigned long flags; + bool enabled; lockdep_assert_held(&prepare_lock); - hlist_for_each_entry(child, &core->children, child_node) - clk_disable_unused_subtree(child); + flags = clk_enable_lock(); - if (core->flags & CLK_OPS_PARENT_ENABLE) - clk_core_prepare_enable(core->parent); + /* Check if the clock is enabled from root to this clock */ + if (parent_enabled) + enabled = clk_core_is_enabled(core); + else + enabled = false; - flags = clk_enable_lock(); + hlist_for_each_entry(child, &core->children, child_node) + /* + * If any child ignored disable, this clock should too, + * unless there is, valid reason for the clock to be enabled + */ + if (clk_disable_unused_subtree(child, enabled) && + enabled && !core->enable_count) + core->flags |= CLK_IGNORE_UNUSED; - if (core->enable_count) + if (core->flags & CLK_IGNORE_UNUSED || core->enable_count) goto unlock_out; - if (core->flags & CLK_IGNORE_UNUSED) + /* + * If the parent is disabled but the gate is open, we should sanitize + * the situation. This will avoid an unexpected enable of the clock as + * soon as the parent is enabled, without control of CCF. + * + * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without + * forcefully enabling a whole part of the subtree. Just let the + * situation resolve it self on the first enable of the clock + */ + if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE)) goto unlock_out; /* @@ -1516,8 +1545,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core) unlock_out: clk_enable_unlock(flags); - if (core->flags & CLK_OPS_PARENT_ENABLE) - clk_core_disable_unprepare(core->parent); + return (core->flags & CLK_IGNORE_UNUSED) && enabled; } static bool clk_ignore_unused __initdata; @@ -1550,16 +1578,16 @@ 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, true); hlist_for_each_entry(core, &clk_orphan_list, child_node) - clk_disable_unused_subtree(core); + clk_disable_unused_subtree(core, true); hlist_for_each_entry(core, &clk_root_list, child_node) - clk_unprepare_unused_subtree(core); + clk_unprepare_unused_subtree(core, true); hlist_for_each_entry(core, &clk_orphan_list, child_node) - clk_unprepare_unused_subtree(core); + clk_unprepare_unused_subtree(core, true); clk_prepare_unlock();
As it as been pointed out numerous times, flagging a clock with CLK_IGNORE_UNUSED does _not_ guarantee that clock left enabled will stay on. The clock will get disabled if any enable/disable cycle happens on it or its parent clocks. Because enable/disable cycles will disable unused clocks, clk_disable_unused() should not trigger such cycle. Doing so disregard the flag if set for any parent clocks. This is problematic with CLK_OPS_PARENT_ENABLE handling. To solve this, and a couple other issues, pass the parent status to the child while walking the subtree, and return whether child ignored disable, or not. * Knowing the parent status allows to safely disable clocks with CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means that, while the clock is not gated it is effectively disabled. Turning on the parents to sanitize the sitation would bring back our initial problem, so just let it sanitize itself when the clock gets used. * If a clock is not actively used (enabled_count == 0), does not have CLK_IGNORE_UNUSED but the hw enabled all the way to the root clock, and a child ignored the disable, it should ignore the disable too. Doing so avoids disabling what is feading the children. Let the flag trickle down the tree. This has the added benefit to transfer the information to the unprepare path, so we don't unprepare the parent of a clock that ignored a disable. * An enabled clock must be prepared in CCF but we can't rely solely on counts at clk_disable_unused() stage. Make sure an enabled clock is considered prepared too, even if does not implement the related callback. Also make sure only disabled clocks get unprepared. Signed-off-by: Jerome Brunet <jbrunet@baylibre.com> --- This is sent as an RFC to continue the discussion started by Chuan. It is not meant to be applied as it is. drivers/clk/clk.c | 92 ++++++++++++++++++++++++++++++----------------- 1 file changed, 60 insertions(+), 32 deletions(-)