On Fri, Nov 08, 2019 at 10:12:49AM -0800, Stephen Boyd wrote:
> Quoting Thierry Reding (2019-11-08 02:11:16)
> > 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.
> 
> Ok cool. Our whole suspend/resume and save/restore story hasn't really
> been well thought out so we may want to pull all this logic into the
> core one day. For now it's OK to do the heavy lifting from provider
> drivers until someone gets a better grasp on how this should all work.

Ah, that would explain why I was scratching my head trying to understand
how exactly this was supposed to work. It did feel like there was some
infrastructure there, but looking around there wasn't a very consistent
usage pattern that I could find.

I think suspend/resume is always a little tricky. For example the clocks
may required a slightly different logical sequences between SoCs. Maybe
even different types of clocks have different needs. We seem to have a
bit of that on Tegra alone already. Without having delved into this too
much, it seems to me like the core can't do a whole lot without stepping
(potentially) on drivers' toes.

The current save_context/restore_context seems to be mostly that,
though, so I think it's a good starting point. You're right that we may
eventually see clearer patterns appear.

> > > 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.
> 
> Right.
> 
> > 
> > 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.
> > 
> 
> Sure a WARN_ON() sounds fair. That will not take the whole task down
> and makes sure that drivers aren't doing something incorrect. Otherwise,
> this looks good and we can optimize by caching the parent index later if
> we really need to.

Okay, great. I'll go replace the above patch in the branch that I have.
I'm not sure if you saw it, but I had sent this in a pull request for
v5.5-rc1 about a week ago because I've got Tegra clock driver patches
that depend on this. I can replace this patch with the above proposal
and update the Tegra clock driver branch and then resend the two pull
requests.

Does that sound like a plan?

Thierry