Message-ID: <87a8jxyfua.fsf@eliezer.anholt.net>
Date:	Tue, 10 May 2016 10:37:17 -0700
From:	Eric Anholt <eric@...olt.net>
To:	Martin Sperl <kernel@...tin.sperl.org>,
	Michael Turquette <mturquette@...libre.com>,
	Stephen Boyd <sboyd@...eaurora.org>
Cc:	linux-kernel@...r.kernel.org, linux-rpi-kernel@...ts.infradead.org,
	linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH 0/3] clk: bcm2835: critical clocks and parent selection

Martin Sperl <kernel@...tin.sperl.org> writes:

> On 10.05.2016 03:01, Eric Anholt wrote:
>> With the new patch 2 inserted between my previous pair, I think this
>> should cover Martin's bugs with clock disabling.
>>
>> I tested patch 2 to be important on the downstream kernel: with the
>> DPI panel support added there, I was losing ethernet (my only I/O)
>> when the HDMI HSM hanging off of PLLD_PER got disabled due to
>> EPROBE_DEFER.
>>
>> Eric Anholt (3):
>>    clk: bcm2835: Mark the VPU clock as critical
>>    clk: bcm2835: Mark GPIO clocks enabled at boot as critical.
>>    clk: bcm2835: Skip PLLC clocks when deciding on a new clock parent
>>
>>   drivers/clk/bcm/clk-bcm2835.c | 32 ++++++++++++++++++++++++++++++--
>>   1 file changed, 30 insertions(+), 2 deletions(-)
>>
> I gave it a try - with all 3 patches applied I get the following enabled 
> clocks:
> root@...pcm:~# grep -vE ^0 /sys/kernel/debug/clk/*/clk_enable_count
> /sys/kernel/debug/clk/aux_uart/clk_enable_count:1
> /sys/kernel/debug/clk/emmc/clk_enable_count:1
> /sys/kernel/debug/clk/gp1/clk_enable_count:1
> /sys/kernel/debug/clk/gp2/clk_enable_count:1
> /sys/kernel/debug/clk/osc/clk_enable_count:1
> /sys/kernel/debug/clk/pllc/clk_enable_count:2
> /sys/kernel/debug/clk/pllc_core0/clk_enable_count:1
> /sys/kernel/debug/clk/pllc_per/clk_enable_count:1
> /sys/kernel/debug/clk/vpu/clk_enable_count:2
>
> At least on my compute module gp1/gp2 is enabled, but there is no rate
> set - so why is it marked as critical for all devices?
> So why apply patch2 for all possible devices?

According to the CLK_IS_CRITICAL patches, the author intended critical
clocks not to use the included function for marking clocks as critical
From the DT.  I'm not sure why, but writing patches using that when they
say not to seemed like a waste.

We could check if gp1/gp2 are already on before marking them critical.

> Loading/unloading the amba_pl011 module does not crash the system,
> but a simple stty -F /dev/ttyAMA0 does crash the system!
>
> Here the sequence:
> root@...pcm:~# dmesg -C
> root@...pcm:~# modprobe amba_pl011
> root@...pcm:~# dmesg -c
> [  141.708453] Serial: AMBA PL011 UART driver
> [  141.709158] 20201000.uart: ttyAMA0 at MMIO 0x20201000 (irq = 81, 
> base_baud = 0) is a PL011 rev2
> root@...pcm:~# rmmod  amba_pl011
> root@...pcm:~# dmesg -c
> [  150.511248] Trying to free nonexistent resource 
> <0000000020201000-0000000020201fff>
> root@...pcm:~# modprobe amba_pl011
> root@...pcm:~# dmesg -c
> [  159.385002] Serial: AMBA PL011 UART driver
> [  159.385714] 20201000.uart: ttyAMA0 at MMIO 0x20201000 (irq = 81, 
> base_baud = 0) is a PL011 rev2
> root@...pcm:~# stty -F /dev/ttyAMA0
> speed 9600 baud; line = 0;
> -brkint -imaxbel
> root@...pcm:~# Timeout, server raspcm not responding.
>
> The reason behind this is that the firmware pre-configured uart clock
> looks like this:
> root@...pcm:~# cat /sys/kernel/debug/clk/uart/regdump
> ctl = 0x00000296
> div = 0x000a6aab
> so it is configured to use plld_per (which itself is running, even if 
> not enabled
> in the kernel)
>
> But as plld_per is not among the enabled clocks then plld_per
> gets disabled as soon as the tty device is closed (by stty) and
> this also disables plld...
>
> Similar effect when using PCM/i2s and use speaker-test:
> root@...pcm:~# dmesg -C
> root@...pcm:~# modprobe snd-soc-bcm2835-i2s; modprobe snd-soc-pcm5102a; 
> modprobe snd-soc-hifiberry-dac
> root@...pcm:~# dmesg
> [   81.968591] snd-hifiberry-dac sound: pcm5102a-hifi <-> 20203000.i2s 
> mapping ok
> root@...pcm:~# speaker-test -c 2 -r 44100 -F S16_LE -f 440 -t sine&
> [1] 579
> root@...pcm:~#
> speaker-test 1.0.28
>
> Playback device is default
> Stream parameters are 44100Hz, S16_LE, 2 channels
> Sine wave rate is 440.0000Hz
> Rate set to 44100Hz (requested 44100Hz)
> Buffer size range from 128 to 131072
> Period size range from 64 to 65536
> Using max buffer size 131072
> Periods = 4
> was set period_size = 32768
> was set buffer_size = 131072
>   0 - Front Left
>   1 - Front Right
>
> root@...pcm:~#
> root@...pcm:~# grep -vE ^0 /sys/kernel/debug/clk/*/clk_enable_count
> /sys/kernel/debug/clk/aux_uart/clk_enable_count:1
> /sys/kernel/debug/clk/emmc/clk_enable_count:1
> /sys/kernel/debug/clk/gp1/clk_enable_count:1
> /sys/kernel/debug/clk/gp2/clk_enable_count:1
> /sys/kernel/debug/clk/osc/clk_enable_count:2
> /sys/kernel/debug/clk/pcm/clk_enable_count:1
> /sys/kernel/debug/clk/pllc/clk_enable_count:2
> /sys/kernel/debug/clk/pllc_core0/clk_enable_count:1
> /sys/kernel/debug/clk/pllc_per/clk_enable_count:1
> /sys/kernel/debug/clk/plld/clk_enable_count:1
> /sys/kernel/debug/clk/plld_per/clk_enable_count:1
> /sys/kernel/debug/clk/vpu/clk_enable_count:2
> root@...pcm:~# kill %1
> root@...pcm:~# Time per period = 106.889502
> Timeout, server raspcm not responding.
>
> You see that plld gets now used and when I kill speaker-test
> the machine crashes again.

Just so I can be clear here: What are you using to talk to the Pi?
Builtin USB ethernet?

> So this patchset does not really solve any of the problems that
> I have reported either.
>
> That is why my patchset has taken the "HAND_OFF" approach
> instead (which still just hides some of the issues), but at least
> it does not crash the system on the use of plld and it allows
> for custom parent and mash selection.

HAND_OFF sure doesn't look like it's landing in the next kernel, so
writing patches using it doesn't make much sense to me.

> In reality it would require consumers of the corresponding
> parent clocks in the kernel (arm, ...) and the knowledge which
> clocks are really needed by the firmware - i.e plld.
>
> Note that the sdram clock is using plld_core parent!
> root@...pcm:~#  cat /sys/kernel/debug/clk/sdram/regdump
> ctl = 0x00004006
> div = 0x00003000
> root@...pcm:~# cat /sys/kernel/debug/clk/sdram/clk_rate
> 166533331

However, it's not enabled, right?  Bit 4 isn't set in the CTL reg.

> and also hsm (probably hardware security module):
> root@...pcm:~# cat /sys/kernel/debug/clk/hsm/regdump
> ctl = 0x000002d6
> div = 0x000030e0
> root@...pcm:~# cat /sys/kernel/debug/clk/hsm/clk_rate
> 163551916

That's the HDMI state machine (there's even a comment saying so),
controlled by the vc4 driver.

> So turning plld off stops sdram and hsm - at least that is my
> interpretation.
>
> This means we need to define a clock property in firmware or
> we need a ram node making use of "mmio-sram" maybe?
>
> Marking sdram as "critical" or "hand_off" could also solve that
> for the moment (but it does not solve all the other hidden
> clock dependencies of the firmware)

If there are other hidden dependencies, then we should figure them out.

> --- a/drivers/clk/bcm/clk-bcm2835.c
> +++ b/drivers/clk/bcm/clk-bcm2835.c
> @@ -1655,7 +1655,8 @@ static const struct bcm2835_clk_desc 
> clk_desc_array[] = {
>                  .ctl_reg = CM_SDCCTL,
>                  .div_reg = CM_SDCDIV,
>                  .int_bits = 6,
> -               .frac_bits = 0),
> +               .frac_bits = 0,
> +               .flags = CLK_IS_CRITICAL),
>          [BCM2835_CLOCK_V3D]     = REGISTER_VPU_CLK(
>                  .name = "v3d",
>                  .ctl_reg = CM_V3DCTL,

The Pi foundation folks believe that the cprman SDRAM clock isn't ever
used (there's a separate PLL in the SDRAM controller, and cprman is only
intended for unused low-power states), and at least in your sample of
the reg, it's not enabled.  Instead of grepping for clk_enable_count, it
would be really useful for debugging to look at your clk_summary
instead.

> Still I would say that this should actually move to the dt to
> correctly describe the HW.

If you created a series that *just* added critical support, using
assigned-clocks, and got the DT folks to agree to it, that would be fine
with me.

Download attachment "signature.asc" of type "application/pgp-signature" (819 bytes)

Powered by blists - more mailing lists