| Message ID | 1438386285-5070-2-git-send-email-sboyd@codeaurora.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Sat, 2015-08-01 at 00:44 +0100, Stephen Boyd wrote: > @@ -179,6 +124,15 @@ static void __init clk_sp810_of_setup(struct device_node *node) > sp810->timerclken[i].channel = i; > sp810->timerclken[i].hw.init = &init; > > + /* > + * If DT isn't setting the parent, force it to be > + * the 1 MHz clock without going through the framework. > + * We do this before clk_register() so that it can determine > + * the parent and setup the tree properly. > + */ > + if (deprecated) > + init.ops->set_parent(&sp810->timerclken[i].hw, 1); > + > sp810->timerclken[i].clk = clk_register(NULL, > &sp810->timerclken[i].hw); > WARN_ON(IS_ERR(sp810->timerclken[i].clk)); So that's one thing I've got a (small) problem here... The above change assumes that SP810 always have 32kHz clock on input 0 and 1MHz clock on input 1. Yes, this is how it made on VExpress, but it doesn't have to be the case. The magic picking up the faster clock was added to handle all possible cases. The bottom line is: if all we care is VExpress than it works, but it's still a hack. Personally I don't like it, however I won't nak the patch because of this. Pawel
On 08/03/2015 07:01 AM, Pawel Moll wrote: > On Sat, 2015-08-01 at 00:44 +0100, Stephen Boyd wrote: >> @@ -179,6 +124,15 @@ static void __init clk_sp810_of_setup(struct device_node *node) >> sp810->timerclken[i].channel = i; >> sp810->timerclken[i].hw.init = &init; >> >> + /* >> + * If DT isn't setting the parent, force it to be >> + * the 1 MHz clock without going through the framework. >> + * We do this before clk_register() so that it can determine >> + * the parent and setup the tree properly. >> + */ >> + if (deprecated) >> + init.ops->set_parent(&sp810->timerclken[i].hw, 1); >> + >> sp810->timerclken[i].clk = clk_register(NULL, >> &sp810->timerclken[i].hw); >> WARN_ON(IS_ERR(sp810->timerclken[i].clk)); > So that's one thing I've got a (small) problem here... > > The above change assumes that SP810 always have 32kHz clock on input 0 > and 1MHz clock on input 1. Yes, this is how it made on VExpress, but it > doesn't have to be the case. The magic picking up the faster clock was > added to handle all possible cases. > > The bottom line is: if all we care is VExpress than it works, but it's > still a hack. Personally I don't like it, however I won't nak the patch > because of this. > All current dts files in the kernel tree have 1MHz on input 1, so we make this change here to set the parent to input 1 if there isn't an assigned-clock-parents property. Presumably new dts files should have the new property so that things work properly. I tried to find the binding document, but it doesn't look to exist, so I didn't have anything to update.
On Mon, 2015-08-03 at 18:55 +0100, Stephen Boyd wrote: > On 08/03/2015 07:01 AM, Pawel Moll wrote: > > On Sat, 2015-08-01 at 00:44 +0100, Stephen Boyd wrote: > >> @@ -179,6 +124,15 @@ static void __init clk_sp810_of_setup(struct device_node *node) > >> sp810->timerclken[i].channel = i; > >> sp810->timerclken[i].hw.init = &init; > >> > >> + /* > >> + * If DT isn't setting the parent, force it to be > >> + * the 1 MHz clock without going through the framework. > >> + * We do this before clk_register() so that it can determine > >> + * the parent and setup the tree properly. > >> + */ > >> + if (deprecated) > >> + init.ops->set_parent(&sp810->timerclken[i].hw, 1); > >> + > >> sp810->timerclken[i].clk = clk_register(NULL, > >> &sp810->timerclken[i].hw); > >> WARN_ON(IS_ERR(sp810->timerclken[i].clk)); > > So that's one thing I've got a (small) problem here... > > > > The above change assumes that SP810 always have 32kHz clock on input 0 > > and 1MHz clock on input 1. Yes, this is how it made on VExpress, but it > > doesn't have to be the case. The magic picking up the faster clock was > > added to handle all possible cases. > > > > The bottom line is: if all we care is VExpress than it works, but it's > > still a hack. Personally I don't like it, however I won't nak the patch > > because of this. > > All current dts files in the kernel tree have 1MHz on input 1, so we > make this change here to set the parent to input 1 if there isn't an > assigned-clock-parents property. That's what I said - all VE platforms known to me will work, because they all have the clocks wired up the same way. And I still don't like code in a "generic SP810 driver" assuming this. Call me what you want ;-) > Presumably new dts files should have > the new property so that things work properly. Sure, no argument here. > I tried to find the > binding document, but it doesn't look to exist, so I didn't have > anything to update. Right, that's probably my fault. The sp810 nodes appeared in the VE trees very early (in terms of the "DT era"), when we weren't so strict about documentation. So far, it would be fairly simple, something like the text below - feel free to take it, mend it, extend it with your changes and include in the series: Thanks! Pawel 8<------------------------- SP810 System Controller ----------------------- Required properties: - compatible: standard compatible string for a Primecell peripheral, see Documentation/devicetree/bindings/arm/primecell.txt for more details should be: "arm,sp810", "arm,primecell" - reg: standard registers property, physical address and size of the control registers - clock-names: from the common clock bindings, for more details see Documentation/devicetree/bindings/clock/clock-bindings.txt; should be: "refclk", "timclk", "apb_pclk" - clocks: from the common clock bindings, phandle and clock specifier pairs for the entries of clock-names property - #clock-cells: from the common clock bindings; should be: <1> - clock-output-names: from the common clock bindings; should be: "timerclken0", "timerclken1", "timerclken2", "timerclken3" Example: sysctl@020000 { compatible = "arm,sp810", "arm,primecell"; reg = <0x020000 0x1000>; clocks = <&v2m_refclk32khz>, <&v2m_refclk1mhz>, <&smbclk>; clock-names = "refclk", "timclk", "apb_pclk"; #clock-cells = <1>; clock-output-names = "timerclken0", "timerclken1", "timerclken2", "timerclken3"; };
On 08/05/2015 03:29 AM, Pawel Moll wrote: > > Right, that's probably my fault. The sp810 nodes appeared in the VE > trees very early (in terms of the "DT era"), when we weren't so strict > about documentation. So far, it would be fairly simple, something like > the text below - feel free to take it, mend it, extend it with your > changes and include in the series: > > Thanks! Great! Can you please provide your signed-off-by? I can make it up into a patch with some commit text that either you or I provide and throw it into the series.
diff --git a/drivers/clk/versatile/clk-sp810.c b/drivers/clk/versatile/clk-sp810.c index 7fbe4d4bf35e..a1cdef6b0f90 100644 --- a/drivers/clk/versatile/clk-sp810.c +++ b/drivers/clk/versatile/clk-sp810.c @@ -33,12 +33,9 @@ struct clk_sp810_timerclken { struct clk_sp810 { struct device_node *node; - int refclk_index, timclk_index; void __iomem *base; spinlock_t lock; struct clk_sp810_timerclken timerclken[4]; - struct clk *refclk; - struct clk *timclk; }; static u8 clk_sp810_timerclken_get_parent(struct clk_hw *hw) @@ -71,55 +68,7 @@ static int clk_sp810_timerclken_set_parent(struct clk_hw *hw, u8 index) return 0; } -/* - * FIXME - setting the parent every time .prepare is invoked is inefficient. - * This is better handled by a dedicated clock tree configuration mechanism at - * init-time. Revisit this later when such a mechanism exists - */ -static int clk_sp810_timerclken_prepare(struct clk_hw *hw) -{ - struct clk_sp810_timerclken *timerclken = to_clk_sp810_timerclken(hw); - struct clk_sp810 *sp810 = timerclken->sp810; - struct clk *old_parent = __clk_get_parent(hw->clk); - struct clk *new_parent; - - if (!sp810->refclk) - sp810->refclk = of_clk_get(sp810->node, sp810->refclk_index); - - if (!sp810->timclk) - sp810->timclk = of_clk_get(sp810->node, sp810->timclk_index); - - if (WARN_ON(IS_ERR(sp810->refclk) || IS_ERR(sp810->timclk))) - return -ENOENT; - - /* Select fastest parent */ - if (clk_get_rate(sp810->refclk) > clk_get_rate(sp810->timclk)) - new_parent = sp810->refclk; - else - new_parent = sp810->timclk; - - /* Switch the parent if necessary */ - if (old_parent != new_parent) { - clk_prepare(new_parent); - clk_set_parent(hw->clk, new_parent); - clk_unprepare(old_parent); - } - - return 0; -} - -static void clk_sp810_timerclken_unprepare(struct clk_hw *hw) -{ - struct clk_sp810_timerclken *timerclken = to_clk_sp810_timerclken(hw); - struct clk_sp810 *sp810 = timerclken->sp810; - - clk_put(sp810->timclk); - clk_put(sp810->refclk); -} - static const struct clk_ops clk_sp810_timerclken_ops = { - .prepare = clk_sp810_timerclken_prepare, - .unprepare = clk_sp810_timerclken_unprepare, .get_parent = clk_sp810_timerclken_get_parent, .set_parent = clk_sp810_timerclken_set_parent, }; @@ -140,24 +89,18 @@ static void __init clk_sp810_of_setup(struct device_node *node) { struct clk_sp810 *sp810 = kzalloc(sizeof(*sp810), GFP_KERNEL); const char *parent_names[2]; + int num = ARRAY_SIZE(parent_names); char name[12]; struct clk_init_data init; int i; + bool deprecated; if (!sp810) { pr_err("Failed to allocate memory for SP810!\n"); return; } - sp810->refclk_index = of_property_match_string(node, "clock-names", - "refclk"); - parent_names[0] = of_clk_get_parent_name(node, sp810->refclk_index); - - sp810->timclk_index = of_property_match_string(node, "clock-names", - "timclk"); - parent_names[1] = of_clk_get_parent_name(node, sp810->timclk_index); - - if (!parent_names[0] || !parent_names[1]) { + if (of_clk_parent_fill(node, parent_names, num) != num) { pr_warn("Failed to obtain parent clocks for SP810!\n"); return; } @@ -170,7 +113,9 @@ static void __init clk_sp810_of_setup(struct device_node *node) init.ops = &clk_sp810_timerclken_ops; init.flags = CLK_IS_BASIC; init.parent_names = parent_names; - init.num_parents = ARRAY_SIZE(parent_names); + init.num_parents = num; + + deprecated = !of_find_property(node, "assigned-clock-parents", NULL); for (i = 0; i < ARRAY_SIZE(sp810->timerclken); i++) { snprintf(name, ARRAY_SIZE(name), "timerclken%d", i); @@ -179,6 +124,15 @@ static void __init clk_sp810_of_setup(struct device_node *node) sp810->timerclken[i].channel = i; sp810->timerclken[i].hw.init = &init; + /* + * If DT isn't setting the parent, force it to be + * the 1 MHz clock without going through the framework. + * We do this before clk_register() so that it can determine + * the parent and setup the tree properly. + */ + if (deprecated) + init.ops->set_parent(&sp810->timerclken[i].hw, 1); + sp810->timerclken[i].clk = clk_register(NULL, &sp810->timerclken[i].hw); WARN_ON(IS_ERR(sp810->timerclken[i].clk));
We're removing struct clk from the clk provider API. This code is calling the consumer APIs to change the parent to a 1 MHz fixed rate clock for each of the clocks that the driver provides. Move to using the assigned-clock-parents DT property for this instead. Because this is an ABI break, detect if the property is missing and fall back to setting the parent explicitly before the clocks are registered. Cc: Pawel Moll <pawel.moll@arm.com> Cc: Linus Walleij <linus.walleij@linaro.org> Cc: Sudeep Holla <sudeep.holla@arm.com> Signed-off-by: Stephen Boyd <sboyd@codeaurora.org> --- drivers/clk/versatile/clk-sp810.c | 76 ++++++++------------------------------- 1 file changed, 15 insertions(+), 61 deletions(-)