Re: [PATCH v5 00/10] clk: implement clock rate protection mechanism

[Jerome Brunet]

On Mon, 2018-04-23 at 11:21 -0700, Michael Turquette wrote:
> Quoting Jerome Brunet (2018-02-02 04:50:28)
> > On Thu, 2018-02-01 at 09:43 -0800, Stephen Boyd wrote:
> > > > > > Applied to clk-protect-rate, with the exception that I did not apply
> > > > > > "clk: fix CLK_SET_RATE_GATE with clock rate protection" as it breaks
> > > > > > qcom clk code.
> > > > > > 
> > > > > > Stephen, do you plan to fix up the qcom clock code so that the
> > > > > > SET_RATE_GATE improvement can go in?
> > > > > > 
> > > > > 
> > > > > I started working on it a while back. Let's see if I can finish
> > > > > it off this weekend.
> > > > > 
> > > > 
> > > > Hi Stephen,
> > > > 
> > > > Have you been able find something to fix the qcom code regarding this issue ?
> > > > 
> > > 
> > > This is what I have. I'm unhappy with a few things. First, I made
> > > a spinlock for each clk, which is overkill. Most likely, just a
> > > single spinlock is needed per clk-controller device. Second, I
> > > haven't finished off the branch/gate part, so gating/ungating of
> > > branches needs to be locked as well to prevent branches from
> > > turning on while rates change. And finally, the 'branches' list is
> > > duplicating a bunch of information about the child clks of an
> > > RCG, so it feels like we need a core framework API to enable and
> > > disable clks forcibly while remembering what is enabled/disabled
> > > or at least to walk the clk tree and call some function.
> > 
> > Looks similar to Mike's CCR idea ;)
> 
> Giving clk provider drivers more control over the clocks that they
> provide is a similar concept, but the ancient ccr series dealt almost
> exclusively with set_rate and set_parent ops. 
> 
> > 
> > > 
> > > The spinlock per clk-controller is duplicating the regmap lock we
> > > already have, so we may want a regmap API to grab the lock, and
> > > then another regmap API to do reads/writes without grabbing the
> > > lock, and then finally release the lock with a regmap unlock API.
> > 
> > There is 'regsequence' for multiple write in a burst, but that's only if you do
> > write only ... I suppose you are more in read/update/writeback mode, so it
> > probably does not help much.
> > 
> > Maybe we could extend regmap's regsequence, to do a sequence of
> > regmap_update_bits() ?
> > 
> > Another possibility could be to provide your own lock/unlock ops when
> > registering the regmap. With this, you might be able to supply your own spinlock
> > to regmap. This is already supported by regmap, would it help ?
> 
> Stephen, was there ever an update on your end? This patch has been
> dangling for a while and I thought it was time to ping on it.
> 
> Regards,
> Mike

Mike, Stephen,
The patch as been sitting on list for 6 months now and CLK_SET_RATE_GATE still
does not really do what it should, at least not completely.
How can we progress on this ?

As explained in this mail [0] from March, I think there is a fairly simple way
to deal with platforms relying on the broken implementation, such as qcom and
the mmci driver.

[0]: https://lkml.kernel.org/r/1522398050.3871.12.camel@xxxxxxxxxxxx

Regards
Jerome

> 
> > 
> > > This part is mostly an optimization, but it would be nice to have
> > > so that multiple writes could be done in sequence. This way, the
> > > RCG code could do the special locking sequence and the branch
> > > code could do the fire and forget single bit update.
> 
>