Re: [PATCH] ARM: OMAP2+: Fix warnings with broken omap2_set_init_voltage()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

* H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> [191202 21:10]:
> > Am 25.09.2019 um 01:32 schrieb Tony Lindgren <tony@xxxxxxxxxxx>:
> > Guys, please check and ack if we can really do this to get rid of some
> > pointless dmesg -l3 errors without affecting the ongoing cpufreq and
> > voltage work.
> 
> unfortunately we did not yet test in combination with the 1GHz OPP
> patches for omap3630 (queued for v5.5-rc1) and it appears that this
> patch breaks the 1GHz OPP.
> 
> The symptom is that it works fine on a dm3730 with 800MHz rating
> but results in spurious kernel panics, stack corruption, virtual memory
> failures, OneNAND DMA timeouts etc. on a dm3730 with 1GHz speed grade.

Hmm yeah OK, I was a bit worried about this breaking something.

> We can also re-add the "turbo-mode" tags in the omap36xx.dtsi (or
> remove the 1GHz OPP) and can then boot with 800MHz max. clock. But
> enabling boost (echo 1 >/sys/devices/system/cpu/cpufreq/boost) makes
> the problem and its symptoms appear almost immediately.
> 
> After some scratching our heads we found that v5.3.7 is still good and
> v5.3.8 is bad. A bisect of our tree (which includes the 1GHz OPP) did
> point to this patch whichwas apparently already backported to v5.3.8 and
> v5.4.
> 
> So I assume that the code removed here enabled or initialized something
> we need for safe 1GHz transitions. Maybe the ABB-LDO. Or it looks up the
> vdd regulator and initializes it earlier than without this code. Maybe
> it also (pre-)initializes some clk which could now be left uninitialized
> too long?

It was just doing voltdm_lookup() and clk_get_rate() and then failed
dev_pm_opp_find_freq_ceil(), but I don't see what these might affect..

> Note that seeing the log message indicates that voltdm_scale() and
> dev_pm_opp_get_voltage() are not called, but all functions before could
> be with side-effects.

Yes that is strange. There's no clk_prepare() before we proceed to
call clk_get_rate() either, not sure if that matter here though.

> v5.5-rc1 will likely fail right from the beginning (only on 1GHz rated
> omap36xx) because it makes the combination of this patch and 1GHz OPP
> public (linux-next should already fail but it appears that nobody has
> tested).

OK

> Any ideas how to fix? Before I try to do a revert and then add goto exit;
> after each function call and see which ones are essential for 1GHz.

If you have things reproducable, care to try to narrow the issue down
a bit by trying see which parts of the old omap2_set_init_voltage()
fix the issue?

The issue should be there somewhere in the few lines of code before
dev_pm_opp_find_freq_ceil(), right?

It would be good to understand what's going on before reverting or
fixing things condering that a revert would add back code that has
it's own errors and fails to init :)

Another thing to check is if the dev instance is actually the right
one we had in omap2_set_init_voltage() vs the dts dev instance as
we use that with dev_pm_opp_find_freq_ceil().

Regards,

Tony



[Index of Archives]     [Linux Arm (vger)]     [ARM Kernel]     [ARM MSM]     [Linux Tegra]     [Linux WPAN Networking]     [Linux Wireless Networking]     [Maemo Users]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux