mbox

[GIT,PULL,4/8] clk: tegra: Changes for v4.2-rc1

Message ID 1431524980-13599-5-git-send-email-thierry.reding@gmail.com
State New
Headers show

Pull-request

git://git.kernel.org/pub/scm/linux/kernel/git/tegra/linux.git tags/tegra-for-4.2-clk

Message

Thierry Reding May 13, 2015, 1:49 p.m. UTC
Hi Mike, Stephen,

The following changes since commit b787f68c36d49bb1d9236f403813641efa74a031:

  Linux 4.1-rc1 (2015-04-26 17:59:10 -0700)

are available in the git repository at:

  git://git.kernel.org/pub/scm/linux/kernel/git/tegra/linux.git tags/tegra-for-4.2-clk

for you to fetch changes up to 36b7be6d3ea8f434f1e0723f3fb0e85c3e00ebc2:

  clk: tegra: Fix hda2codec_2x clock name for Tegra30 (2015-05-13 15:17:14 +0200)

I've based this pull request on top of the tegra-for-4.2-ramcode pull
request, so pulling only this one should be sufficient to resolve the
dependency.

Thanks,
Thierry

----------------------------------------------------------------
clk: tegra: Changes for v4.2-rc1

This contains the EMC clock driver that's been exhaustively reviewed and
tested. It also includes a change to the clock core that allows a clock
provider to perform low-level reparenting of clocks. This is required by
the EMC clock driver because the reparenting needs to be done at a very
specific point in time during the EMC frequency switch.

----------------------------------------------------------------
Marcel Ziswiler (1):
      clk: tegra: Fix hda2codec_2x clock name for Tegra30

Mikko Perttunen (3):
      soc/tegra: fuse: Add RAM code reader helper
      clk: tegra: Remove old Tegra124 EMC clock
      clk: tegra: Add EMC clock driver

Thierry Reding (2):
      Merge branch 'for-4.2/ramcode' into for-4.2/clk
      clk: tegra: EMC clock driver depends on EMC driver

Tomeu Vizoso (6):
      of: Document long-ram-code property in nvidia,tegra20-apbmisc
      clk: Expose clk_hw_reparent() to providers
      of: document new emc-timings subnode in nvidia,tegra124-car
      of: document external-memory-controller property in tegra124-car
      clk: tegra: Set the EMC clock as the parent of the MC clock
      clk: tegra: Have EMC clock implement determine_rate()

 .../bindings/clock/nvidia,tegra124-car.txt         |  44 +-
 .../bindings/misc/nvidia,tegra20-apbmisc.txt       |   2 +
 drivers/clk/Kconfig                                |   1 +
 drivers/clk/clk.c                                  |   8 +
 drivers/clk/tegra/Kconfig                          |   3 +
 drivers/clk/tegra/Makefile                         |   1 +
 drivers/clk/tegra/clk-emc.c                        | 538 +++++++++++++++++++++
 drivers/clk/tegra/clk-tegra124.c                   |  19 +-
 drivers/clk/tegra/clk-tegra30.c                    |   2 +-
 drivers/clk/tegra/clk.h                            |  12 +
 drivers/soc/tegra/fuse/tegra-apbmisc.c             |  21 +
 include/linux/clk-provider.h                       |   1 +
 include/soc/tegra/fuse.h                           |   1 +
 13 files changed, 637 insertions(+), 16 deletions(-)
 create mode 100644 drivers/clk/tegra/Kconfig
 create mode 100644 drivers/clk/tegra/clk-emc.c

Comments

Stephen Boyd May 20, 2015, 7:54 p.m. UTC | #1
On 05/13, Thierry Reding wrote:
> Hi Mike, Stephen,
> 
> The following changes since commit b787f68c36d49bb1d9236f403813641efa74a031:
> 
>   Linux 4.1-rc1 (2015-04-26 17:59:10 -0700)
> 
> are available in the git repository at:
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/tegra/linux.git tags/tegra-for-4.2-clk
> 
> for you to fetch changes up to 36b7be6d3ea8f434f1e0723f3fb0e85c3e00ebc2:
> 
>   clk: tegra: Fix hda2codec_2x clock name for Tegra30 (2015-05-13 15:17:14 +0200)
> 
> I've based this pull request on top of the tegra-for-4.2-ramcode pull
> request, so pulling only this one should be sufficient to resolve the
> dependency.
> 
> Thanks,
> Thierry
> 
> ----------------------------------------------------------------
> clk: tegra: Changes for v4.2-rc1
> 
> This contains the EMC clock driver that's been exhaustively reviewed and
> tested. It also includes a change to the clock core that allows a clock
> provider to perform low-level reparenting of clocks. This is required by
> the EMC clock driver because the reparenting needs to be done at a very
> specific point in time during the EMC frequency switch.

Can someone please describe why we need to do software
reparenting at a specific point in time during a frequency
switch? I must have missed out on the conversation somewhere and
looking at the commit that introduces the function, the argument
for why the API is exposed:

      To be used by clock implementations for switching to a new
      parent during rate change.

is lacking in any details about *why* we need it.
Mikko Perttunen May 21, 2015, 6:25 a.m. UTC | #2
On 05/20/2015 10:54 PM, Stephen Boyd wrote:
> On 05/13, Thierry Reding wrote:
>> Hi Mike, Stephen,
>>
>> The following changes since commit b787f68c36d49bb1d9236f403813641efa74a031:
>>
>>   Linux 4.1-rc1 (2015-04-26 17:59:10 -0700)
>>
>> are available in the git repository at:
>>
>>   git://git.kernel.org/pub/scm/linux/kernel/git/tegra/linux.git tags/tegra-for-4.2-clk
>>
>> for you to fetch changes up to 36b7be6d3ea8f434f1e0723f3fb0e85c3e00ebc2:
>>
>>   clk: tegra: Fix hda2codec_2x clock name for Tegra30 (2015-05-13 15:17:14 +0200)
>>
>> I've based this pull request on top of the tegra-for-4.2-ramcode pull
>> request, so pulling only this one should be sufficient to resolve the
>> dependency.
>>
>> Thanks,
>> Thierry
>>
>> ----------------------------------------------------------------
>> clk: tegra: Changes for v4.2-rc1
>>
>> This contains the EMC clock driver that's been exhaustively reviewed and
>> tested. It also includes a change to the clock core that allows a clock
>> provider to perform low-level reparenting of clocks. This is required by
>> the EMC clock driver because the reparenting needs to be done at a very
>> specific point in time during the EMC frequency switch.
> 
> Can someone please describe why we need to do software
> reparenting at a specific point in time during a frequency
> switch? I must have missed out on the conversation somewhere and
> looking at the commit that introduces the function, the argument
> for why the API is exposed:
> 
>       To be used by clock implementations for switching to a new
>       parent during rate change.
> 
> is lacking in any details about *why* we need it.
> 

Hi Stephen,

the way the EMC clock rate is set in hardware is similar to other
clocks, i.e. there's a register and you write the new divider and parent
id to it. However, with EMC, you cannot just do this any time you want;
writing to the register initiates a state machine in the clock hardware
that changes a large number of other parameters regarding DRAM timings
as well. These parameters need to be programmed into shadow registers
before the rate change write is done. This means that both the new
divisor and the parent must be written in the same register write.

The CCF has a callback, set_rate_and_parent, which allows for both to be
passed to the driver at the same time. However, it also requires
set_rate and set_parent to be implemented, which the EMC driver cannot do.

Furthermore, the CCF cannot help with parent selection for the EMC clock
at all. The parent for each rate is selected by the board designer.

There is also the issue that sometimes, the EMC driver cannot directly
switch to the target (rate, parent) pair; instead it is necessary to
switch first to another pair we call a backup timing. In this situation,
we temporarily have a parent that is neither the one we had before the
set_rate call or the one it will be after. Especially, if the switch to
the backup timing succeeds but the following switch to the target timing
fails, then without the reparent call the parent in hardware would be
left inconsistent with the software state.

This is why we've decided to implement the driver with only the set_rate
callback and decide the required transitions within the driver.
This also means that the driver needs to have a function to tell the CCF
that it has changed its parent.

Thanks,
Mikko.

nvpublic
Mikko Perttunen May 27, 2015, 2:59 p.m. UTC | #3
On 05/21/2015 09:25 AM, Mikko Perttunen wrote:
> On 05/20/2015 10:54 PM, Stephen Boyd wrote:
>> On 05/13, Thierry Reding wrote:
>>> ...

Hi Mike, Stephen,

just reminding about this PR :)

Thanks,
Mikko.
Stephen Boyd May 27, 2015, 7:50 p.m. UTC | #4
On 05/21, Mikko Perttunen wrote:
> On 05/20/2015 10:54 PM, Stephen Boyd wrote:
> > On 05/13, Thierry Reding wrote:
> >> Hi Mike, Stephen,
> >>
> >> The following changes since commit b787f68c36d49bb1d9236f403813641efa74a031:
> >>
> >>   Linux 4.1-rc1 (2015-04-26 17:59:10 -0700)
> >>
> >> are available in the git repository at:
> >>
> >>   git://git.kernel.org/pub/scm/linux/kernel/git/tegra/linux.git tags/tegra-for-4.2-clk
> >>
> >> for you to fetch changes up to 36b7be6d3ea8f434f1e0723f3fb0e85c3e00ebc2:
> >>
> >>   clk: tegra: Fix hda2codec_2x clock name for Tegra30 (2015-05-13 15:17:14 +0200)
> >>
> >> I've based this pull request on top of the tegra-for-4.2-ramcode pull
> >> request, so pulling only this one should be sufficient to resolve the
> >> dependency.
> >>
> >> Thanks,
> >> Thierry
> >>
> >> ----------------------------------------------------------------
> >> clk: tegra: Changes for v4.2-rc1
> >>
> >> This contains the EMC clock driver that's been exhaustively reviewed and
> >> tested. It also includes a change to the clock core that allows a clock
> >> provider to perform low-level reparenting of clocks. This is required by
> >> the EMC clock driver because the reparenting needs to be done at a very
> >> specific point in time during the EMC frequency switch.
> > 
> > Can someone please describe why we need to do software
> > reparenting at a specific point in time during a frequency
> > switch? I must have missed out on the conversation somewhere and
> > looking at the commit that introduces the function, the argument
> > for why the API is exposed:
> > 
> >       To be used by clock implementations for switching to a new
> >       parent during rate change.
> > 
> > is lacking in any details about *why* we need it.
> > 
> 
> Hi Stephen,
> 
> the way the EMC clock rate is set in hardware is similar to other
> clocks, i.e. there's a register and you write the new divider and parent
> id to it. However, with EMC, you cannot just do this any time you want;
> writing to the register initiates a state machine in the clock hardware
> that changes a large number of other parameters regarding DRAM timings
> as well. These parameters need to be programmed into shadow registers
> before the rate change write is done. This means that both the new
> divisor and the parent must be written in the same register write.
> 
> The CCF has a callback, set_rate_and_parent, which allows for both to be
> passed to the driver at the same time. However, it also requires
> set_rate and set_parent to be implemented, which the EMC driver cannot do.

Ok so we should change the framework to allow drivers to only
implement set_rate_and_parent and use that in set_rate and
set_parent implementations if it's the only option available. Or
is there a problem if CCF allows clk_set_parent() on the EMC
clock by calling set_rate_and_parent() with the new parent and
whatever recalc_rate returns for the new parent's rate going into
the clock?

> 
> Furthermore, the CCF cannot help with parent selection for the EMC clock
> at all. The parent for each rate is selected by the board designer.

I'm not following this part. The CCF mostly asks clock providers
what parent should be used when clk_set_rate() is called.

> 
> There is also the issue that sometimes, the EMC driver cannot directly
> switch to the target (rate, parent) pair; instead it is necessary to
> switch first to another pair we call a backup timing. In this situation,
> we temporarily have a parent that is neither the one we had before the
> set_rate call or the one it will be after. Especially, if the switch to
> the backup timing succeeds but the following switch to the target timing
> fails, then without the reparent call the parent in hardware would be
> left inconsistent with the software state.

Yes, I've sent a patch a while ago to support an idea of a "safe
parent" or a backup timing that can be used while a clock is
reprogrammed. Mike has also proposed the concept of "coordinated
clock rates" which would do something similar and allow clock
providers to control complicated rate transitions themselves. It
seems that we may be able to use the "safe parent" concept here,
where first we switch to some other parent, call the set_rate*
op, and then switch the parent back if we're not already using
the parent that we should be using.

This sort of thing becomes important for things like per-clock
locking where we need to know how the clock tree is going to
change *before* we modify the clock tree. If we go back and forth
between the clock providers and the clock tree while we're in the
middle of making irreversible changes to the hardware, we may get
stuck in a situation where we need to lock more clocks and then
we may need to undo hardware changes.
Mike Turquette June 18, 2015, 11:41 p.m. UTC | #5
Quoting Mikko Perttunen (2015-05-28 00:03:01)
> On 05/27/2015 10:50 PM, Stephen Boyd wrote:
> > On 05/21, Mikko Perttunen wrote:
> >>
> >> Hi Stephen,
> >>
> >> the way the EMC clock rate is set in hardware is similar to other
> >> clocks, i.e. there's a register and you write the new divider and parent
> >> id to it. However, with EMC, you cannot just do this any time you want;
> >> writing to the register initiates a state machine in the clock hardware
> >> that changes a large number of other parameters regarding DRAM timings
> >> as well. These parameters need to be programmed into shadow registers
> >> before the rate change write is done. This means that both the new
> >> divisor and the parent must be written in the same register write.
> >>
> >> The CCF has a callback, set_rate_and_parent, which allows for both to be
> >> passed to the driver at the same time. However, it also requires
> >> set_rate and set_parent to be implemented, which the EMC driver cannot do.
> > 
> > Ok so we should change the framework to allow drivers to only
> > implement set_rate_and_parent and use that in set_rate and
> > set_parent implementations if it's the only option available. Or
> > is there a problem if CCF allows clk_set_parent() on the EMC
> > clock by calling set_rate_and_parent() with the new parent and
> > whatever recalc_rate returns for the new parent's rate going into
> > the clock?
> 
> There isn't really a problem, but the EMC driver cannot implement this
> operation sensibly so it would just always return an error if the (rate,
> parent) pair given to set_rate_and_parent() doesn't exactly match one of
> the entries specified in device tree.
> 
> > 
> >>
> >> Furthermore, the CCF cannot help with parent selection for the EMC clock
> >> at all. The parent for each rate is selected by the board designer.
> > 
> > I'm not following this part. The CCF mostly asks clock providers
> > what parent should be used when clk_set_rate() is called.
> 
> Yep, that much is fine; what I was alluding was the above (set_parent
> and set_rate_and_parent with an unlisted (rate, parent) pair are invalid)
> 
> > 
> >>
> >> There is also the issue that sometimes, the EMC driver cannot directly
> >> switch to the target (rate, parent) pair; instead it is necessary to
> >> switch first to another pair we call a backup timing. In this situation,
> >> we temporarily have a parent that is neither the one we had before the
> >> set_rate call or the one it will be after. Especially, if the switch to
> >> the backup timing succeeds but the following switch to the target timing
> >> fails, then without the reparent call the parent in hardware would be
> >> left inconsistent with the software state.
> > 
> > Yes, I've sent a patch a while ago to support an idea of a "safe
> > parent" or a backup timing that can be used while a clock is
> > reprogrammed. Mike has also proposed the concept of "coordinated
> > clock rates" which would do something similar and allow clock
> > providers to control complicated rate transitions themselves. It
> > seems that we may be able to use the "safe parent" concept here,
> > where first we switch to some other parent, call the set_rate*
> > op, and then switch the parent back if we're not already using
> > the parent that we should be using.
> 
> While I'm not sure how much sophistication is warranted in the CCF, for
> our use case this backup timing has to be able to be a function of the
> timing we are leaving and preferably also the target timing. Also, as
> usual, the backup timings are (rate, parent) pairs, and just changing
> rate or just changing parent are both impossible.
> 
> > 
> > This sort of thing becomes important for things like per-clock
> > locking where we need to know how the clock tree is going to
> > change *before* we modify the clock tree. If we go back and forth
> > between the clock providers and the clock tree while we're in the
> > middle of making irreversible changes to the hardware, we may get
> > stuck in a situation where we need to lock more clocks and then
> > we may need to undo hardware changes.
> > 
> 
> Yeah, makes sense.
> 
> Do you think you can still pull this patchset? There's other code in
> linux-next that depends on it and I'd prefer to have a working driver in
> the kernel; if/when the improvements to CCF materialize we could update
> the driver to use them.

I'm not really sure what to do with this PR. This seems to fall into the
same category as the Exynos "cpu clocks" series: you have a complex
sequence that requires multiple clock nodes to be changes in a
coordinated fashion.

I'm working on some core infrastructure to fix this. I'd like to get
that on the list asap and possibly merged for 4.3. I think it can
benefit your case and remove the need to export clk_hw_reparent, which
is pretty nasty.

What exactly will break if this is not pulled? I appreciate your offer
to update this driver when the core changes are merged, but I would
prefer to do it the right way first, instead of fixing up something that
is already merged after the fact.

Regards,
Mike

> 
> Thanks,
> Mikko.
Mike Turquette June 20, 2015, 8:41 p.m. UTC | #6
Quoting Michael Turquette (2015-06-18 16:41:26)
> Quoting Mikko Perttunen (2015-05-28 00:03:01)
> > On 05/27/2015 10:50 PM, Stephen Boyd wrote:
> > > On 05/21, Mikko Perttunen wrote:
> > >>
> > >> Hi Stephen,
> > >>
> > >> the way the EMC clock rate is set in hardware is similar to other
> > >> clocks, i.e. there's a register and you write the new divider and parent
> > >> id to it. However, with EMC, you cannot just do this any time you want;
> > >> writing to the register initiates a state machine in the clock hardware
> > >> that changes a large number of other parameters regarding DRAM timings
> > >> as well. These parameters need to be programmed into shadow registers
> > >> before the rate change write is done. This means that both the new
> > >> divisor and the parent must be written in the same register write.
> > >>
> > >> The CCF has a callback, set_rate_and_parent, which allows for both to be
> > >> passed to the driver at the same time. However, it also requires
> > >> set_rate and set_parent to be implemented, which the EMC driver cannot do.
> > > 
> > > Ok so we should change the framework to allow drivers to only
> > > implement set_rate_and_parent and use that in set_rate and
> > > set_parent implementations if it's the only option available. Or
> > > is there a problem if CCF allows clk_set_parent() on the EMC
> > > clock by calling set_rate_and_parent() with the new parent and
> > > whatever recalc_rate returns for the new parent's rate going into
> > > the clock?
> > 
> > There isn't really a problem, but the EMC driver cannot implement this
> > operation sensibly so it would just always return an error if the (rate,
> > parent) pair given to set_rate_and_parent() doesn't exactly match one of
> > the entries specified in device tree.
> > 
> > > 
> > >>
> > >> Furthermore, the CCF cannot help with parent selection for the EMC clock
> > >> at all. The parent for each rate is selected by the board designer.
> > > 
> > > I'm not following this part. The CCF mostly asks clock providers
> > > what parent should be used when clk_set_rate() is called.
> > 
> > Yep, that much is fine; what I was alluding was the above (set_parent
> > and set_rate_and_parent with an unlisted (rate, parent) pair are invalid)
> > 
> > > 
> > >>
> > >> There is also the issue that sometimes, the EMC driver cannot directly
> > >> switch to the target (rate, parent) pair; instead it is necessary to
> > >> switch first to another pair we call a backup timing. In this situation,
> > >> we temporarily have a parent that is neither the one we had before the
> > >> set_rate call or the one it will be after. Especially, if the switch to
> > >> the backup timing succeeds but the following switch to the target timing
> > >> fails, then without the reparent call the parent in hardware would be
> > >> left inconsistent with the software state.
> > > 
> > > Yes, I've sent a patch a while ago to support an idea of a "safe
> > > parent" or a backup timing that can be used while a clock is
> > > reprogrammed. Mike has also proposed the concept of "coordinated
> > > clock rates" which would do something similar and allow clock
> > > providers to control complicated rate transitions themselves. It
> > > seems that we may be able to use the "safe parent" concept here,
> > > where first we switch to some other parent, call the set_rate*
> > > op, and then switch the parent back if we're not already using
> > > the parent that we should be using.
> > 
> > While I'm not sure how much sophistication is warranted in the CCF, for
> > our use case this backup timing has to be able to be a function of the
> > timing we are leaving and preferably also the target timing. Also, as
> > usual, the backup timings are (rate, parent) pairs, and just changing
> > rate or just changing parent are both impossible.
> > 
> > > 
> > > This sort of thing becomes important for things like per-clock
> > > locking where we need to know how the clock tree is going to
> > > change *before* we modify the clock tree. If we go back and forth
> > > between the clock providers and the clock tree while we're in the
> > > middle of making irreversible changes to the hardware, we may get
> > > stuck in a situation where we need to lock more clocks and then
> > > we may need to undo hardware changes.
> > > 
> > 
> > Yeah, makes sense.
> > 
> > Do you think you can still pull this patchset? There's other code in
> > linux-next that depends on it and I'd prefer to have a working driver in
> > the kernel; if/when the improvements to CCF materialize we could update
> > the driver to use them.
> 
> I'm not really sure what to do with this PR. This seems to fall into the
> same category as the Exynos "cpu clocks" series: you have a complex
> sequence that requires multiple clock nodes to be changes in a
> coordinated fashion.

I've decided to pull this in the interest of fairness. I'm taking the
Exynos cpu clock patches, which will need to be updated in the future to
use some new infrastructure for coordinating rate changes across
multiple clock nodes. I'm merging the EMC stuff with the same
expectation that it will need to migrate to the new infrastructure when
it becomes available and we'll get rid of clk_hw_reparent.

Sound good?

Regards,
Mike

> 
> I'm working on some core infrastructure to fix this. I'd like to get
> that on the list asap and possibly merged for 4.3. I think it can
> benefit your case and remove the need to export clk_hw_reparent, which
> is pretty nasty.
> 
> What exactly will break if this is not pulled? I appreciate your offer
> to update this driver when the core changes are merged, but I would
> prefer to do it the right way first, instead of fixing up something that
> is already merged after the fact.
> 
> Regards,
> Mike
> 
> > 
> > Thanks,
> > Mikko.
Mikko Perttunen June 22, 2015, 6:59 a.m. UTC | #7
On 06/20/15 23:41, Michael Turquette wrote:
> Quoting Michael Turquette (2015-06-18 16:41:26)
>> Quoting Mikko Perttunen (2015-05-28 00:03:01)
>>> On 05/27/2015 10:50 PM, Stephen Boyd wrote:
>>>> On 05/21, Mikko Perttunen wrote:
>>>>>
>>>>> Hi Stephen,
>>>>>
>>>>> the way the EMC clock rate is set in hardware is similar to other
>>>>> clocks, i.e. there's a register and you write the new divider and parent
>>>>> id to it. However, with EMC, you cannot just do this any time you want;
>>>>> writing to the register initiates a state machine in the clock hardware
>>>>> that changes a large number of other parameters regarding DRAM timings
>>>>> as well. These parameters need to be programmed into shadow registers
>>>>> before the rate change write is done. This means that both the new
>>>>> divisor and the parent must be written in the same register write.
>>>>>
>>>>> The CCF has a callback, set_rate_and_parent, which allows for both to be
>>>>> passed to the driver at the same time. However, it also requires
>>>>> set_rate and set_parent to be implemented, which the EMC driver cannot do.
>>>>
>>>> Ok so we should change the framework to allow drivers to only
>>>> implement set_rate_and_parent and use that in set_rate and
>>>> set_parent implementations if it's the only option available. Or
>>>> is there a problem if CCF allows clk_set_parent() on the EMC
>>>> clock by calling set_rate_and_parent() with the new parent and
>>>> whatever recalc_rate returns for the new parent's rate going into
>>>> the clock?
>>>
>>> There isn't really a problem, but the EMC driver cannot implement this
>>> operation sensibly so it would just always return an error if the (rate,
>>> parent) pair given to set_rate_and_parent() doesn't exactly match one of
>>> the entries specified in device tree.
>>>
>>>>
>>>>>
>>>>> Furthermore, the CCF cannot help with parent selection for the EMC clock
>>>>> at all. The parent for each rate is selected by the board designer.
>>>>
>>>> I'm not following this part. The CCF mostly asks clock providers
>>>> what parent should be used when clk_set_rate() is called.
>>>
>>> Yep, that much is fine; what I was alluding was the above (set_parent
>>> and set_rate_and_parent with an unlisted (rate, parent) pair are invalid)
>>>
>>>>
>>>>>
>>>>> There is also the issue that sometimes, the EMC driver cannot directly
>>>>> switch to the target (rate, parent) pair; instead it is necessary to
>>>>> switch first to another pair we call a backup timing. In this situation,
>>>>> we temporarily have a parent that is neither the one we had before the
>>>>> set_rate call or the one it will be after. Especially, if the switch to
>>>>> the backup timing succeeds but the following switch to the target timing
>>>>> fails, then without the reparent call the parent in hardware would be
>>>>> left inconsistent with the software state.
>>>>
>>>> Yes, I've sent a patch a while ago to support an idea of a "safe
>>>> parent" or a backup timing that can be used while a clock is
>>>> reprogrammed. Mike has also proposed the concept of "coordinated
>>>> clock rates" which would do something similar and allow clock
>>>> providers to control complicated rate transitions themselves. It
>>>> seems that we may be able to use the "safe parent" concept here,
>>>> where first we switch to some other parent, call the set_rate*
>>>> op, and then switch the parent back if we're not already using
>>>> the parent that we should be using.
>>>
>>> While I'm not sure how much sophistication is warranted in the CCF, for
>>> our use case this backup timing has to be able to be a function of the
>>> timing we are leaving and preferably also the target timing. Also, as
>>> usual, the backup timings are (rate, parent) pairs, and just changing
>>> rate or just changing parent are both impossible.
>>>
>>>>
>>>> This sort of thing becomes important for things like per-clock
>>>> locking where we need to know how the clock tree is going to
>>>> change *before* we modify the clock tree. If we go back and forth
>>>> between the clock providers and the clock tree while we're in the
>>>> middle of making irreversible changes to the hardware, we may get
>>>> stuck in a situation where we need to lock more clocks and then
>>>> we may need to undo hardware changes.
>>>>
>>>
>>> Yeah, makes sense.
>>>
>>> Do you think you can still pull this patchset? There's other code in
>>> linux-next that depends on it and I'd prefer to have a working driver in
>>> the kernel; if/when the improvements to CCF materialize we could update
>>> the driver to use them.
>>
>> I'm not really sure what to do with this PR. This seems to fall into the
>> same category as the Exynos "cpu clocks" series: you have a complex
>> sequence that requires multiple clock nodes to be changes in a
>> coordinated fashion.
>
> I've decided to pull this in the interest of fairness. I'm taking the
> Exynos cpu clock patches, which will need to be updated in the future to
> use some new infrastructure for coordinating rate changes across
> multiple clock nodes. I'm merging the EMC stuff with the same
> expectation that it will need to migrate to the new infrastructure when
> it becomes available and we'll get rid of clk_hw_reparent.
>
> Sound good?

Sounds good! Please keep us informed and once the framework changes land 
we'll fix up this one.

>
> Regards,
> Mike
>

Thanks,
Mikko.

>>
>> I'm working on some core infrastructure to fix this. I'd like to get
>> that on the list asap and possibly merged for 4.3. I think it can
>> benefit your case and remove the need to export clk_hw_reparent, which
>> is pretty nasty.
>>
>> What exactly will break if this is not pulled? I appreciate your offer
>> to update this driver when the core changes are merged, but I would
>> prefer to do it the right way first, instead of fixing up something that
>> is already merged after the fact.
>>
>> Regards,
>> Mike
>>
>>>
>>> Thanks,
>>> Mikko.
> --
> To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
>
Mikko Perttunen June 22, 2015, 7:03 a.m. UTC | #8
Fixing Mike's mail address

On 06/22/15 09:59, Mikko Perttunen wrote:
> On 06/20/15 23:41, Michael Turquette wrote:
>> Quoting Michael Turquette (2015-06-18 16:41:26)
>> ...
>> I've decided to pull this in the interest of fairness. I'm taking the
>> Exynos cpu clock patches, which will need to be updated in the future to
>> use some new infrastructure for coordinating rate changes across
>> multiple clock nodes. I'm merging the EMC stuff with the same
>> expectation that it will need to migrate to the new infrastructure when
>> it becomes available and we'll get rid of clk_hw_reparent.
>>
>> Sound good?
>
> Sounds good! Please keep us informed and once the framework changes land
> we'll fix up this one.
>
>>
>> Regards,
>> Mike
>>
>
> Thanks,
> Mikko.
>
>>>
>>> I'm working on some core infrastructure to fix this. I'd like to get
>>> that on the list asap and possibly merged for 4.3. I think it can
>>> benefit your case and remove the need to export clk_hw_reparent, which
>>> is pretty nasty.
>>>
>>> What exactly will break if this is not pulled? I appreciate your offer
>>> to update this driver when the core changes are merged, but I would
>>> prefer to do it the right way first, instead of fixing up something that
>>> is already merged after the fact.
>>>
>>> Regards,
>>> Mike
>>>
>>>>
>>>> Thanks,
>>>> Mikko.
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
>>
>
Thierry Reding June 29, 2015, 8:54 a.m. UTC | #9
On Sat, Jun 20, 2015 at 01:41:04PM -0700, Michael Turquette wrote:
[...]
> I've decided to pull this in the interest of fairness. I'm taking the
> Exynos cpu clock patches, which will need to be updated in the future to
> use some new infrastructure for coordinating rate changes across
> multiple clock nodes. I'm merging the EMC stuff with the same
> expectation that it will need to migrate to the new infrastructure when
> it becomes available and we'll get rid of clk_hw_reparent.
> 
> Sound good?

Sounds good, thanks Mike. I'm sure we'll find someone to update this
when the framework changes are ready.

Thierry