On Thu, Nov 07, 2019 at 11:19:32AM -0800, Stephen Boyd wrote:
> Quoting Thierry Reding (2019-11-07 07:21:15)
> > On Thu, Nov 07, 2019 at 03:54:03AM +0300, Dmitry Osipenko wrote:
> > > 07.11.2019 02:10, Stephen Boyd пишет:
> > > > Quoting Sowjanya Komatineni (2019-08-16 12:41:52)
> > > >> This patch adds an API clk_hw_get_parent_index to get index of the
> > > >> clock parent to use during the clock restore operations on system
> > > >> resume.
> > > >  
> > > > Is there a reason we can't save the clk hw index at suspend time by
> > > > reading the hardware to understand the current parent? The parent index
> > > > typically doesn't matter unless we're trying to communicate something
> > > > from the framework to the provider driver. Put another way, I would
> > > > think the provider driver can figure out the index itself without having
> > > > to go through the framework to do so.
> > > 
> > > Isn't it a bit wasteful to duplicate information about the parent within
> > > a provider if framework already has that info? The whole point of this
> > > new API is to allow providers to avoid that unnecessary duplication.
> > > 
> > > Please note that clk_hw_get_parent_index is getting used only at the
> > > resume time and not at suspend.
> > 
> > I agree with this. All of the information that we need is already cached
> > in the framework. Doing this in the driver would mean essentially adding
> > a "saved parent" field along with code to read the value at suspend time
> > to the three types of clocks that currently use this core helper.
> 
> Don't we already have a "saved parent" field by storing the pointer to
> the clk_hw?
> 
> > 
> > That's certainly something that we *can* do, but it doesn't sound like a
> > better option than simply querying the framework for the value that we
> > need.
> > 
> 
> Let me say this another way. Why does this driver want to know the index
> that the framework uses for some clk_hw pointer? Perhaps it happens to
> align with the same value that hardware uses, but I still don't
> understand why the driver wants to know what the framework has decided
> is the index for some clk_hw pointer.
> 
> Or is this something like "give me the index for the parent that the
> framework thinks I currently have but in reality don't have anymore
> because the register contents were wiped and we need to reparent it"?

Yeah, that's exactly what this is being used for. It's used to restore
the parent/child relationship during resume after the registers have
been wiped during supend.

> A generic API to get any index for this question is overkill and we should
> consider adding some sort of API like clk_hw_get_current_parent_index(),
> or a framework flag that tells the framework this parent is incorrect
> and we need to call the .set_parent() op again to reconfigure it.

Okay, I think I see what you're saying. The current implementation does
carry a bit of a risk because users could be calling this function with
any arbitrary pair of struct clk_hw *, even completely unrelated ones.

How about we turn it into this instead:

	/**
	 * clk_hw_get_parent_index - return the index of the parent clock
	 * @hw: clk_hw associated with the clk being consumed
	 *
	 * Fetches and returns the index of parent clock. Returns -EINVAL if the given
	 * clock does not have a current parent.
	 */
	int clk_hw_get_parent_index(struct clk_hw *hw)
	{
		struct clk_hw *parent = clk_hw_get_parent(hw);

		if (!parent)
			return -EINVAL;

		return clk_fetch_parent_index(hw->core, parent->core);
	}
	EXPORT_SYMBOL_GPL(clk_hw_get_parent_index);

I think that has the advantage that we can't pass it a parent that's not
really a parent. There's still the slightly weird case where the clock
doesn't have a current parent, but hopefully that's something we are not
going to encounter much. After all this only makes sense to be called on
mux clocks and they always do have a parent by definition.

Perhaps we should be more explicit and wrap that !parent conditional in
a WARN_ON()? In my local patches I do that at the call sites because
they are all functions returning void, so we'd be silently ignoring the
cases, but I think it may make sense to have it in the core.

Any thoughts?

Thierry