mbox series

[v2,0/6] Meson8b: make the CPU clock mutable

Message ID 20181115224048.13511-1-martin.blumenstingl@googlemail.com (mailing list archive)
Headers show
Series Meson8b: make the CPU clock mutable | expand

Message

Martin Blumenstingl Nov. 15, 2018, 10:40 p.m. UTC
This allows changing the CPU clock on the 32-bit Amlogic Meson SoCs
(Meson8, Meson8b and Meson8m2).
CPU frequency scaling will be enabled with a separate series by adding
the CPU clock and the OPP tables to meson8.dtsi and meson8b.dtsi.

While changing the CPU frequency (sys_pll or any of it's post-dividers)
we need to run the CPU clock off the XTAL clock. Otherwise the system
will lock up because we need to disable the sys_pll to change it's
rate.

This also makes the clk-pll's .enable hook a no-op if the clock is
already enabled. Otherwise we're getting lockups when calling the
first clk_{prepare_}enable on the sys_pll or any of it's children (as
the CCF propagates the enable event up to the sys_pll). This is because
the .enable hook unconditionally disables and enables the clock.
However, we can't disable that clock (not even temporarily) if the CPU
is running off sys_pll.

Additionally this adds support for more M/N combinations in sys_pll to
achieve all of the OPPs on Meson8b and all OPPs <= 1608 MHz on Meson8
and Meson8m2.

Compared to Amlogic's 3.10 kernel there's one notable difference: we
are actually allowing changes to the sys_pll. Amlogic's kernel sets
sys_pll to a fixed rate during boot and then uses a timer generate a
"virtual clock rate" by toggling between various dividers (for example:
sys_pll is set to 1536MHz. to achieve 1008MHz they are toggling every
2500us between 1536MHZ and 768MHz so the average over <period, for
example one second> is 1008MHz).
I could reproduce any situation where changing sys_pll failed (for
example due to high temperature). To prove that I ran "stress --cpu 4"
for multiple hours and then cycled through all available CPU
frequencies (while keeping "stress" running in the background). This
worked fine on my Meson8b Odroid-C1 and EC-100 boards as well as my
Meson8m2 board.


Dependencies:
This series is built on top of the latest clk-meson.git tree and the
patches from my other series [1] "Meson8b: fixes for the cpu_scale_div
clock".


Changes since v1 at [0]:
- re-ordered patches as suggested by Jerome: keep all fixes first, then
  the new "features"
- squashed patches "clk-pll: check if the clock is already enabled" and
  "clk-pll: add the is_enabled function in the clk_ops". This also
  allows calling clk_hw_is_enabled() from meson_clk_pll_enable()
  instead of calling meson_clk_pll_is_enabled() directly (this matches
  the implementation of sclk-div.c)
- documented the dependencies of this series in the cover-letter
- dropped "RFC" prefix
- collected Jerome's Acked-/Reviewed-by's (thanks for the quick
  response!)


[0] https://patchwork.kernel.org/cover/10683317/
[1] https://patchwork.kernel.org/cover/10617617/


Martin Blumenstingl (6):
  clk: meson: clk-pll: check if the clock is already enabled
  clk: meson: meson8b: do not use cpu_div3 for cpu_scale_out_sel
  clk: meson: meson8b: mark the CPU clock as CLK_IS_CRITICAL
  clk: meson: meson8b: add support for more M/N values in sys_pll
  clk: meson: meson8b: run from the XTAL when changing the CPU frequency
  clk: meson: meson8b: allow changing the CPU clock tree

 drivers/clk/meson/clk-pll.c | 19 ++++++++
 drivers/clk/meson/meson8b.c | 94 +++++++++++++++++++++++++++++++++----
 2 files changed, 104 insertions(+), 9 deletions(-)

Comments

Neil Armstrong Nov. 16, 2018, 8:49 a.m. UTC | #1
On 15/11/2018 23:40, Martin Blumenstingl wrote:
> This allows changing the CPU clock on the 32-bit Amlogic Meson SoCs
> (Meson8, Meson8b and Meson8m2).
> CPU frequency scaling will be enabled with a separate series by adding
> the CPU clock and the OPP tables to meson8.dtsi and meson8b.dtsi.
> 
> While changing the CPU frequency (sys_pll or any of it's post-dividers)
> we need to run the CPU clock off the XTAL clock. Otherwise the system
> will lock up because we need to disable the sys_pll to change it's
> rate.
> 
> This also makes the clk-pll's .enable hook a no-op if the clock is
> already enabled. Otherwise we're getting lockups when calling the
> first clk_{prepare_}enable on the sys_pll or any of it's children (as
> the CCF propagates the enable event up to the sys_pll). This is because
> the .enable hook unconditionally disables and enables the clock.
> However, we can't disable that clock (not even temporarily) if the CPU
> is running off sys_pll.
> 
> Additionally this adds support for more M/N combinations in sys_pll to
> achieve all of the OPPs on Meson8b and all OPPs <= 1608 MHz on Meson8
> and Meson8m2.
> 
> Compared to Amlogic's 3.10 kernel there's one notable difference: we
> are actually allowing changes to the sys_pll. Amlogic's kernel sets
> sys_pll to a fixed rate during boot and then uses a timer generate a
> "virtual clock rate" by toggling between various dividers (for example:
> sys_pll is set to 1536MHz. to achieve 1008MHz they are toggling every
> 2500us between 1536MHZ and 768MHz so the average over <period, for
> example one second> is 1008MHz).
> I could reproduce any situation where changing sys_pll failed (for
> example due to high temperature). To prove that I ran "stress --cpu 4"
> for multiple hours and then cycled through all available CPU
> frequencies (while keeping "stress" running in the background). This
> worked fine on my Meson8b Odroid-C1 and EC-100 boards as well as my
> Meson8m2 board.
> 
> 
> Dependencies:
> This series is built on top of the latest clk-meson.git tree and the
> patches from my other series [1] "Meson8b: fixes for the cpu_scale_div
> clock".
> 
> 
> Changes since v1 at [0]:
> - re-ordered patches as suggested by Jerome: keep all fixes first, then
>   the new "features"
> - squashed patches "clk-pll: check if the clock is already enabled" and
>   "clk-pll: add the is_enabled function in the clk_ops". This also
>   allows calling clk_hw_is_enabled() from meson_clk_pll_enable()
>   instead of calling meson_clk_pll_is_enabled() directly (this matches
>   the implementation of sclk-div.c)
> - documented the dependencies of this series in the cover-letter
> - dropped "RFC" prefix
> - collected Jerome's Acked-/Reviewed-by's (thanks for the quick
>   response!)
> 
> 
> [0] https://patchwork.kernel.org/cover/10683317/
> [1] https://patchwork.kernel.org/cover/10617617/
> 
> 
> Martin Blumenstingl (6):
>   clk: meson: clk-pll: check if the clock is already enabled
>   clk: meson: meson8b: do not use cpu_div3 for cpu_scale_out_sel
>   clk: meson: meson8b: mark the CPU clock as CLK_IS_CRITICAL
>   clk: meson: meson8b: add support for more M/N values in sys_pll
>   clk: meson: meson8b: run from the XTAL when changing the CPU frequency
>   clk: meson: meson8b: allow changing the CPU clock tree
> 
>  drivers/clk/meson/clk-pll.c | 19 ++++++++
>  drivers/clk/meson/meson8b.c | 94 +++++++++++++++++++++++++++++++++----
>  2 files changed, 104 insertions(+), 9 deletions(-)
> 

Applied to next/drivers !

Thanks for your work on meson8* SoCs !

Neil