Re: [PATCH V5 11/18] clk: tegra210: Add support for Tegra210 clocks

From: Sowjanya Komatineni
Date: Wed Jul 17 2019 - 14:51:46 EST


On 7/17/19 11:32 AM, Dmitry Osipenko wrote:
17.07.2019 20:29, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/17/19 8:17 AM, Dmitry Osipenko wrote:
17.07.2019 9:36, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 11:33 PM, Dmitry Osipenko wrote:
Ð Tue, 16 Jul 2019 22:55:52 -0700
Sowjanya Komatineni <skomatineni@xxxxxxxxxx> ÐÐÑÐÑ:

On 7/16/19 10:42 PM, Dmitry Osipenko wrote:
Ð Tue, 16 Jul 2019 22:25:25 -0700
Sowjanya Komatineni <skomatineni@xxxxxxxxxx> ÐÐÑÐÑ:
On 7/16/19 9:11 PM, Dmitry Osipenko wrote:
Ð Tue, 16 Jul 2019 19:35:49 -0700
Sowjanya Komatineni <skomatineni@xxxxxxxxxx> ÐÐÑÐÑ:
On 7/16/19 7:18 PM, Sowjanya Komatineni wrote:
On 7/16/19 3:06 PM, Sowjanya Komatineni wrote:
On 7/16/19 3:00 PM, Dmitry Osipenko wrote:
17.07.2019 0:35, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 2:21 PM, Dmitry Osipenko wrote:
17.07.2019 0:12, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 1:47 PM, Dmitry Osipenko wrote:
16.07.2019 22:26, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 11:43 AM, Dmitry Osipenko wrote:
16.07.2019 21:30, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 11:25 AM, Dmitry Osipenko wrote:
16.07.2019 21:19, Sowjanya Komatineni ÐÐÑÐÑ:
On 7/16/19 9:50 AM, Sowjanya Komatineni wrote:
On 7/16/19 8:00 AM, Dmitry Osipenko wrote:
16.07.2019 11:06, Peter De Schrijver ÐÐÑÐÑ:
On Tue, Jul 16, 2019 at 03:24:26PM +0800, Joseph
Lo wrote:
OK, Will add to CPUFreq driver...
The other thing that also need attention is
that T124 CPUFreq
driver
implicitly relies on DFLL driver to be probed
first, which is
icky.
Should I add check for successful dfll clk
register explicitly in
CPUFreq driver probe and defer till dfll clk
registers?
Probably you should use the "device links". See
[1][2] for the
example.

[1]
https://elixir.bootlin.com/linux/v5.2.1/source/drivers/gpu/drm/tegra/dc.c#L2383









[2]
https://www.kernel.org/doc/html/latest/driver-api/device_link.html





Return EPROBE_DEFER instead of EINVAL if
device_link_add() fails.
And
use of_find_device_by_node() to get the DFLL's
device, see [3].

[3]
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/drivers/devfreq/tegra20-devfreq.c#n100







Will go thru and add...
Looks like I initially confused this case with getting
orphaned clock.
I'm now seeing that the DFLL driver registers the
clock and then
clk_get(dfll) should be returning EPROBE_DEFER until
DFLL driver is
probed, hence everything should be fine as-is and
there is no real
need
for the 'device link'. Sorry for the confusion!
Sorry, I didn't follow the mail thread. Just
regarding the DFLL
part.

As you know it, the DFLL clock is one of the CPU
clock sources and
integrated with DVFS control logic with the
regulator. We will not
switch
CPU to other clock sources once we switched to
DFLL. Because the
CPU has
been regulated by the DFLL HW with the DVFS table
(CVB or OPP
table
you see
in the driver.). We shouldn't reparent it to
other sources with
unknew
freq/volt pair. That's not guaranteed to work. We
allow switching to
open-loop mode but different sources.
Okay, then the CPUFreq driver will have to enforce
DFLL freq to
PLLP's
rate before switching to PLLP in order to have a
proper CPU voltage.
PLLP freq is safe to work for any CPU voltage. So no
need to enforce
DFLL freq to PLLP rate before changing CCLK_G source
to PLLP during
suspend
Sorry, please ignore my above comment. During
suspend, need to change
CCLK_G source to PLLP when dfll is in closed loop
mode first and
then
dfll need to be set to open loop.
Okay.
And I don't exactly understand why we need to
switch to PLLP in
CPU
idle
driver. Just keep it on CL-DVFS mode all the
time.

In SC7 entry, the dfll suspend function moves it
the open-loop
mode. That's
all. The sc7-entryfirmware will handle the rest
of the sequence to
turn off
the CPU power.

In SC7 resume, the warmboot code will handle the
sequence to
turn on
regulator and power up the CPU cluster. And leave
it on PLL_P.
After
resuming to the kernel, we re-init DFLL, restore
the CPU clock
policy (CPU
runs on DFLL open-loop mode) and then moving to
close-loop mode.
The DFLL is re-inited after switching CCLK to DFLL
parent during of
the
early clocks-state restoring by CaR driver. Hence
instead of having
odd
hacks in the CaR driver, it is much nicer to have a
proper suspend-resume sequencing of the device
drivers. In this case
CPUFreq
driver is the driver that enables DFLL and switches
CPU to that
clock
source, which means that this driver is also should
be responsible for
management of the DFLL's state during of
suspend/resume process. If
CPUFreq driver disables DFLL during suspend and
re-enables it
during
resume, then looks like the CaR driver hacks around
DFLL are not
needed.
The DFLL part looks good to me. BTW, change the
patch subject to
"Add
suspend-resume support" seems more appropriate to
me.
To clarify this, the sequences for DFLL use are as
follows (assuming
all
required DFLL hw configuration has been done)

Switch to DFLL:
0) Save current parent and frequency
1) Program DFLL to open loop mode
2) Enable DFLL
3) Change cclk_g parent to DFLL
For OVR regulator:
4) Change PWM output pin from tristate to output
5) Enable DFLL PWM output
For I2C regulator:
4) Enable DFLL I2C output
6) Program DFLL to closed loop mode

Switch away from DFLL:
0) Change cclk_g parent to PLLP so the CPU
frequency is ok for
any
vdd_cpu voltage
1) Program DFLL to open loop mode
I see during switch away from DFLL (suspend), cclk_g
parent is not
changed to PLLP before changing dfll to open loop
mode.

Will add this ...
The CPUFreq driver switches parent to PLLP during the
probe, similar
should be done on suspend.

I'm also wondering if it's always safe to switch to
PLLP in the probe.
If CPU is running on a lower freq than PLLP, then some
other more
appropriate intermediate parent should be selected.
CPU parents are PLL_X, PLL_P, and dfll. PLL_X always
runs at higher
rate
so switching to PLL_P during CPUFreq probe prior to
dfll clock enable
should be safe.
AFAIK, PLLX could run at ~200MHz. There is also a
divided output of
PLLP
which CCLKG supports, the PLLP_OUT4.

Probably, realistically, CPU is always running off a
fast PLLX during
boot, but I'm wondering what may happen on KEXEC. I
guess ideally CPUFreq driver should also have a
'shutdown' callback to teardown DFLL
on a reboot, but likely that there are other
clock-related problems as
well that may break KEXEC and thus it is not very
important at the
moment.

[snip]
During bootup CPUG sources from PLL_X. By PLL_P source
above I meant
PLL_P_OUT4.

As per clock policies, PLL_X is always used for high freq
like
800Mhz
and for low frequency it will be sourced from PLLP.
Alright, then please don't forget to pre-initialize
PLLP_OUT4 rate to a
reasonable value using tegra_clk_init_table or
assigned-clocks.
PLLP_OUT4 rate update is not needed as it is safe to run at
408Mhz because it is below fmax @ Vmin
So even 204MHz CVB entries are having the same voltage as
408MHz, correct? It's not instantly obvious to me from the
DFLL driver's code where the fmax @ Vmin is defined, I see
that there is the min_millivolts
and frequency entries starting from 204MHZ defined
per-table.
Yes at Vmin CPU Fmax is ~800Mhz. So anything below that will
work at Vmin voltage and PLLP max is 408Mhz.
Thank you for the clarification. It would be good to have that
commented
in the code as well.
OK, Will add...
Regarding, adding suspend/resume to CPUFreq, CPUFreq suspend
happens very early even before disabling non-boot CPUs and also
need to export clock driver APIs to CPUFreq.

Was thinking of below way of implementing this...


Clock DFLL driver Suspend:

ÂÂÂ ÂÂÂ ÂÂÂ - Save CPU clock policy registers, and Perform dfll
suspend which sets in open loop mode

CPU Freq driver Suspend: does nothing


Clock DFLL driver Resume:

ÂÂÂ ÂÂÂ ÂÂÂ - Re-init DFLL, Set in Open-Loop mode, restore CPU
Clock policy registers which actually sets source to DFLL along
with other CPU Policy register restore.

CPU Freq driver Resume:

ÂÂÂ ÂÂÂ ÂÂÂ - do clk_prepare_enable which acutally sets DFLL in
Closed loop mode


Adding one more note: Switching CPU Clock to PLLP is not needed
as CPU CLock can be from dfll in open-loop mode as DFLL is not
disabled anywhere throught the suspend/resume path and SC7 entry
FW and Warm boot code will switch CPU source to PLLP.
Since CPU resumes on PLLP, it will be cleaner to suspend it on
PLLP as well. And besides, seems that currently disabling DFLL
clock will disable DFLL completely and then you'd want to re-init
the DFLL on resume any ways. So better to just disable DFLL
completely on suspend, which should happen on clk_disable(dfll).
Will switch to PLLP during CPUFreq suspend. With decision of using
clk_disable during suspend, its mandatory to switch to PLLP as DFLL
is completely disabled.

My earlier concern was on restoring CPU policy as we can't do that
from CPUFreq driver and need export from clock driver.

Clear now and will do CPU clock policy restore in after dfll
re-init.
Why the policy can't be saved/restored by the CaR driver as a
context of any other clock?
restoring cpu clock policy involves programming source and
super_cclkg_divider.

cclk_g is registered as clk_super_mux and it doesn't use frac_div ops
to do save/restore its divider.
That can be changed of course and I guess it also could be as simple as
saving and restoring of two raw u32 values of the policy/divider
registers.

Also, during clock context we cant restore cclk_g as cclk_g source
will be dfll and dfll will not be resumed/re-initialized by the time
clk_super_mux save/restore happens.

we can't use save/restore context for dfll clk_ops because
dfllCPU_out parent to CCLK_G is first in the clock tree and dfll_ref
and dfll_soc peripheral clocks are not restored by the time dfll
restore happens. Also dfll peripheral clock enables need to be
restored before dfll restore happens which involves programming dfll
controller for re-initialization.

So dfll resume/re-init is done in clk-tegra210 at end of all clocks
restore in V5 series but instead of in clk-tegra210 driver I moved
now to dfll-fcpu driver pm_ops as all dfll dependencies will be
restored thru clk_restore_context by then. This will be in V6.
Since DFLL is now guaranteed to be disabled across CaR suspend/resume
(hence it has nothing to do in regards to CCLK) and given that PLLs
state is restored before the rest of the clocks, I don't see why not to
implement CCLK save/restore in a generic fasion. CPU policy wull be
restored to either PLLP or PLLX (if CPUFreq driver is disabled).

CCLK_G save/restore should happen in clk_super_mux ops save/context and
clk_super_mux save/restore happens very early as cclk_g is first in the
clock tree and save/restore traverses through the tree top-bottom order.
If CCLK_G is restored before the PLLs, then just change the clocks order
such that it won't happen.

I dont think we can change clocks order for CCLK_G.

During bootup, cclk_g is registered after all pll's and peripheral
clocks which is the way we wanted, So cclk_g will be the first one in
the clk list as clk_register adds new clock first in the list.

When clk_save_context and clk_restore_context APIs iterates over the
list, cclk_g is the first
Looking at clk_core_restore_context(), I see that it walks up CLKs list
from parent to children, hence I don't understand how it can ever happen
that CCLK will be restored before the parent. The clocks registration
order doesn't matter at all in that case.

yes from parent to children and dfllCPU_out is the top in the list and its child is cclk_g.

the way clocks are registered is the order I see in the clock list and looking into clk_register API it adds new node first in the list.

cclkg_g & dfll register happens after all plls and peripheral clocks as it need ref, soc and peripheral clocks to be enabled.

So they are the last to get registered and so becomes first in the list.

During save/restore context, it traverses thru this list and first in the list is dfllcpu_OUT (parent) and its child (cclk_g)

saving should not be an issue at all but we cant restore cclk_g/dfll in normal way thru clk_ops restore as plls and peripherals restore doesn't happen by that time.

DFLL enable thru CPUFreq resume happens after all clk_restore_context
happens. So during clk_restore_context, dfll re-init doesnt happen and
doing cpu clock policy restore during super_mux clk_ops will crash as
DFLL is not initialized and its clock is not enabled but CPU clock
restore sets source to DFLL if we restore during super_clk_mux
If CPU was suspended on PLLP, then it will be restored on PLLP by CaR. I
don't understand what DFLL has to do with the CCLK in that case during
the clocks restore.
My above comment is in reference to your request of doing save/restore
for cclk_g in normal fashion thru save/restore context. Because of the
clk order I mentioned above, we cclk_g will be the first one to go thru
save/context.

During save_context of cclk_g, source can be from PLLX, dfll.

Issue will be when we do restore during clk_restore_context of cclk_g as
by that time PLLX/dfll will not be restored.

Seems we already agreed that DFLL will be disabled by the CPUFreq driver
on suspend. Hence CCLK can't be from DFLL if CPU is reparented to PLLP
on CPUFreq driver's suspend, otherwise CPU keeps running from a
boot-state PLLX if CPUFreq driver is disabled.

Yes suspend should not be an issue but issue will be during resume where if we do cclk_g restore in normal way thru clk_restore_context, cclk_g restore happens very early as dfllCPU out is the first one that goes thru restore context and plls/peripherals are not resumed by then.

CPU runs from PLLX if dfll clock enable fails during boot. So when it gets to suspend, we save CPU running clock source as either PLLX or DFLL and then we switch to PLLP.


On resume, CPU runs from PLLP by warm boot code and we need to restore back its source to the one it was using from saved source context (which can be either PLLX or DFLL)

So PLLs & DFLL resume need to happen before CCLKG restore/resume.


With all above discussions, we do DFLL disable in CPUFreq driver on suspend and on CPUFreq resume we enable DFLL back and restore CPU clock source it was using during suspend (which will be either PLLX if dfll enable fails during probe or it will be using DFLL).

So i was trying to say dfll/cclk_g restore can't be done in normal way thru clk_ops save/restore context