Re: [PATCH v13 5/5] drivers: qcom: Update rpmh clients to use start and end transactions

[Doug Anderson]

Hi,

On Tue, Mar 10, 2020 at 4:47 AM Maulik Shah <mkshah@xxxxxxxxxxxxxx> wrote:
>
>
> On 3/10/2020 5:14 AM, Doug Anderson wrote:
> > Hi,
> >
> > On Mon, Mar 9, 2020 at 2:31 AM Maulik Shah <mkshah@xxxxxxxxxxxxxx> wrote:
> >> Update all rpmh clients to start using rpmh_start_transaction() and
> >> rpmh_end_transaction().
> >>
> >> Cc: Taniya Das <tdas@xxxxxxxxxxxxxx>
> >> Cc: Odelu Kukatla <okukatla@xxxxxxxxxxxxxx>
> >> Cc: Kiran Gunda <kgunda@xxxxxxxxxxxxxx>
> >> Cc: Sibi Sankar <sibis@xxxxxxxxxxxxxx>
> >> Signed-off-by: Maulik Shah <mkshah@xxxxxxxxxxxxxx>
> >> ---
> >>  drivers/clk/qcom/clk-rpmh.c             | 21 ++++++++++++++-------
> >>  drivers/interconnect/qcom/bcm-voter.c   | 13 +++++++++----
> >>  drivers/regulator/qcom-rpmh-regulator.c |  4 ++++
> >>  drivers/soc/qcom/rpmhpd.c               | 11 +++++++++--
> > This needs to be 4 separate patches since the change to each subsystem
> > will go through a different maintainer.
> I will split to 4 changes, and send each one to its respective mailing lists and maintainer/reviewer.
> >
> > Also: it'll be a lot easier to land this if you make the new
> > rpmh_start_transaction() and rpmh_end_transaction() calls _optional_
> > for now, especially since they are just a speed optimization and not
> > for correctness.  That is, if a driver makes a call to rpmh_write(),
> > rpmh_write_async(), rpmh_write_batch(), or rpmh_invalidate() without
> > doing rpmh_start_transaction() then it should still work
>
> yes, this is already taken care.
>
> All the calls from driver will go through as it is and won't fail even without calling new APIs.
> So they are already optional.
>
> The comment in rpmh_start_transaction() is already saying if client "choose" to invoke this
> then this must be ended by calling rpmh_end_transaction(). if client don't invoke
> rpmh_start_transaction() in the first place then everything is expected work as if no change.

I think you may have misunderstood.  With your v13 it's not optional
because the flush won't happen unless the clients call
rpmh_start_transaction() and rpmh_end_transaction().  ...but the flush
is necessary for correctness, right?


> > --just flush
> > right away.
>
> No, currently also in driver no one is calling rpmh_flush().
>
> so nothing breaks with series and no point in adding changes to flush right away and then remove them in same series.
>
> when the clients starts invoking new APIs, rpmh_flush() will start getting invoked for the first time in driver.

I'm not saying to expose flush to clients.  I'm saying that you should
modify the implementation of the calls in rpmh.c.  AKA in rpmh.c you
have:

rpmh_write():
  start_transaction()
  ... the contents of rpmh_write() from your v13 ...
  end_transaction()

rpmh_write_async():
  start_transaction()
  ... the contents of rpmh_write_async() from your v13 ...
  end_transaction()

rpmh_write_batch()
  start_transaction()
  ... the contents of rpmh_write_batch() from your v13 ...
  end_transaction()

rpmh_invalidate()
  start_transaction()
  ... the contents of rpmh_invalidate() from your v13 ...
  end_transaction()


If a client calls rpmh_write() without anything else, they will get an
implicit flush.

If a client does this:

start_transaction()
rpmh_invalidate()
rpmh_write_batch()
rpmh_write_batch()
end_transaction()

That translates to:

start_transaction()
  start_transaction()
  ... the contents of rpmh_invalidate() from your v13 ...
  end_transaction()
  start_transaction()
  ... the contents of rpmh_write_batch() from your v13 ...
  end_transaction()
  start_transaction()
  ... the contents of rpmh_write_batch() from your v13 ...
  end_transaction()
end_transaction()

...then you'll get one flush at the end.  That's because the
start_transaction() embedded in rpmh_invalidate() and
rpmh_write_batch() will increase the usage count to 2 and then
decrease back to 1.  Since it won't be 0 you won't get flushes from
the embedded start/end.  You'll just get one flush at the end.


Now start_transaction() and end_transaction() are truly optional.  If
a client doesn't call them they'll get an implicit flush at the end of
every call.  If they do call start/end themselves they can postpone
the flush till the true end.

...and because it's truly optional, you can drop several of the
changes in your series.



-Doug