Re: [PATCH 1/3] clocksource: davinci: work around a clocksource problem on dm365 SoC

From: Bartosz Golaszewski <brgl@bgdev.pl>
To: Sekhar Nori <nsekhar@ti.com>
Cc: Daniel Lezcano <daniel.lezcano@linaro.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	David Lechner <david@lechnology.com>,
	Kevin Hilman <khilman@kernel.org>,
	Linux ARM <linux-arm-kernel@lists.infradead.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Bartosz Golaszewski <bgolaszewski@baylibre.com>
Subject: Re: [PATCH 1/3] clocksource: davinci: work around a clocksource problem on dm365 SoC
Date: Thu, 19 Dec 2019 09:59:47 +0100	[thread overview]
Message-ID: <CAMRc=MdBxko3PGK8MksiOYbm081oQf6gYxBp8xOwtyTJ5ruM5Q@mail.gmail.com> (raw)
In-Reply-To: <51eb10e9-245e-7b3e-51ff-578e06e0759b@ti.com>

śr., 18 gru 2019 o 10:28 Sekhar Nori <nsekhar@ti.com> napisał(a):
>
> Hi Bart,
>
> On 13/12/19 9:54 PM, Bartosz Golaszewski wrote:
> > From: Bartosz Golaszewski <bgolaszewski@baylibre.com>
> >
> > The DM365 platform has a strange quirk (only present when using ancient
> > u-boot - mainline u-boot v2013.01 and later works fine) where if we
> > enable the second half of the timer in periodic mode before we do its
> > initialization - the time won't start flowing and we can't boot.
> >
> > When using more recent u-boot, we can enable the timer, then reinitialize
> > it and all works fine.
> >
> > I've been unable to figure out why that is, but a workaround for this
> > is straightforward - just cache the enable bits for tim34.
> >
> > Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>
>
> Timer Global Control Register (TGCR) has bits to reset both halves of
> timer. Does placing both halves in reset, waiting a bit (say 10ms) and
> then taking them out of reset help solve the this problem?
>

No, it doesn't change anything. On u-boot present on my dm365-evm,
tim34 is not in reset when we get to linux, while tim12 is in reset,
but putting tim34 in reset in linux doesn't seem to change anything.

> Also, there are LPSCs controlling the timers. As an experiment, can you
> see if using LPSC_STATE_SWRSTDISABLE instead of LPSC_STATE_DISABLE in
> davinci_lpsc_clk_disable() and then doing a clk_disable() + clk_enable()
> on timer can get the timer out of this bad state.

I tried several combinations of this e.g. normal prepare_enable ->
disable -> enable, disable -> enable, disable -> delay -> enable etc.
and neither worked.

>
> We need some way for Linux to start on a clean state after bootloader is
> done. And trying to reset the timer before use seems to be a better way
> to accomplish it.
>
> I assume the original code was just lucky in not hitting this case?
>

I guess so. It used to re-read the registers instead of assuming
certain values. When I did it too, there was no problem, it's only
when we dropped re-reading that this must have appeared.

Bart