[PATCH v2 1/5] clocksource: mediatek: do not enable GPT_CLK_EVT when setup

[Yingjoe Chen]

On Thu, 2015-08-20 at 16:28 +0200, Daniel Lezcano wrote:
> On 08/17/2015 04:10 PM, Yingjoe Chen wrote:
> > On Thu, 2015-08-13 at 10:35 +0200, Daniel Lezcano wrote:
> >> On 07/22/2015 10:14 AM, Yingjoe Chen wrote:
> >>> Spurious mtk timer interrupt is noticed at boot and cause kernel
> >>> crash. It seems if GPT is enabled, it will latch irq status even
> >>> when its IRQ is disabled.
> 
> It "seems" ?

Hi,

Datasheet doesn't mention detail. So I did some experiments, playing
around with registers. Based on my observation, I think this is what
happens:

For each GPT timer, it has ENABLE, IRQ_EN, IRQ status, IRQ_ACK, counter
& compare.

When mtk_timer_init calls mtk_timer_setup to setup GPT_CLK_EVT, it
enable the timer but didn't set counter or compare. Both counter &
compare is zero on reset, so GPT immediately raise IRQ status. IRQ_EN is
still disabled now, so it didn't trigger interrupt right away.

At end of mtk_timer_init, it calls mtk_timer_enable_irq to enable irq.
Since IRQ status is 1 now, GPT trigger interrupt immediately. The
interrupt is serviced by mtk_timer_interrupt. Since this is not an
expected event, evt->dev.event_handler will be NULL and system crashed
in the handler.

> >>> When irq is enabled afterward, we see
> >>> spurious interrupt.
> 
> Doesn't have the firmware something to do with that ?

We have 6 GPT on mt8173, mtk timer use 2 of them. The spurious interrupt
only happens on GPT_CLK_EVT (GPT1). Our firmware didn't touch that one,
so it is in reset default when mtk timer driver try to enable it.

> I have a mtk 8173 board I can use next week. How do you reproduce the 
> issue ?
>
> >>> Change init flow to only enable GPT_CLK_SRC at mtk_timer_init.
> >>>
> >>> Acked-by: Matthias Brugger <matthias.bgg at gmail.com>
> >>> Reviewed-by: Daniel Kurtz <djkurtz at chromium.org>
> >>> Signed-off-by: Yingjoe Chen <yingjoe.chen at mediatek.com>
> >>> ---
> >>>
> >>> Update to my patch [1], added __init as Daniel suggest. This is the
> >>> only patch that need to change in that series, so I only sent this one.
> >>>
> >>> http://lists.infradead.org/pipermail/linux-mediatek/2015-July/001545.html
> >>>
> >>>    drivers/clocksource/mtk_timer.c | 16 ++++++++++------
> >>>    1 file changed, 10 insertions(+), 6 deletions(-)
> >>>
> >>> diff --git a/drivers/clocksource/mtk_timer.c b/drivers/clocksource/mtk_timer.c
> >>> index 68ab423..2ba5b66 100644
> >>> --- a/drivers/clocksource/mtk_timer.c
> >>> +++ b/drivers/clocksource/mtk_timer.c
> >>> @@ -156,9 +156,11 @@ static void mtk_timer_global_reset(struct mtk_clock_event_device *evt)
> >>>    	writel(0x3f, evt->gpt_base + GPT_IRQ_ACK_REG);
> >>>    }
> >>>
> >>> -static void
> >>> -mtk_timer_setup(struct mtk_clock_event_device *evt, u8 timer, u8 option)
> >>> +static void __init mtk_timer_setup(struct mtk_clock_event_device *evt,
> >>> +				   u8 timer, u8 option, bool enable)
> >>>    {
> >>> +	u32 val;
> >>> +
> >>>    	writel(TIMER_CTRL_CLEAR | TIMER_CTRL_DISABLE,
> >>>    		evt->gpt_base + TIMER_CTRL_REG(timer));
> >>>
> >>> @@ -167,8 +169,10 @@ mtk_timer_setup(struct mtk_clock_event_device *evt, u8 timer, u8 option)
> >>>
> >>>    	writel(0x0, evt->gpt_base + TIMER_CMP_REG(timer));
> >>>
> >>> -	writel(TIMER_CTRL_OP(option) | TIMER_CTRL_ENABLE,
> >>> -			evt->gpt_base + TIMER_CTRL_REG(timer));
> >>> +	val = TIMER_CTRL_OP(option);
> >>> +	if (enable)
> >>> +		val |= TIMER_CTRL_ENABLE;
> >>> +	writel(val, evt->gpt_base + TIMER_CTRL_REG(timer));
> >>
> >> Instead of the 'enable' new option, I prefer a test with 'timer' with a
> >> comment:
> >>
> >> 	/*
> >> 	 * the timer hw is broken in that way ... bla bla, so we only
> >> 	 * enable the clocksource ...
> >> 	 */
> >> 	if (timer == GPT_CLK_SRC)
> >> 		val |= TIMER_CTRL_ENABLE;
> >
> > Hi Daniel,
> >
> > Thanks for your review.
> > Since this bug happens to anyone using interrupt,
> 
> Can you elaborate ? I don't get the point.
> 
> > I'm not sure checking
> > timer and only enable it for GPT_CLK_SRC is easier to read. Anyway, I'll
> > change to this in next version.
> >> That said, can you have a look at commit 1096be08 ?
> >> "clockevents: sun5i: Fix setup_irq init sequence"
> >>
> >> first and check if moving the interrupt request after the
> >> clockevents_config_and_register could fix your issue.
> >
> > I've tested this before, see:
> >
> > http://lists.infradead.org/pipermail/linux-mediatek/2015-May/000539.html
> > http://lists.infradead.org/pipermail/linux-mediatek/2015-May/000551.html
> 
> I could take or ack this patch trusting it fixes the issue but there are 
> some points that need clarifications.
> 
>   - Does the spurious interrupt occurs *every* time ? at each boot ?

Yes. If you applied this series to enable mtk timer without this fix on
mt8173 or mt8135 you can reproduce it. It occurs for every boot.

It crash before uart driver is ready, so you'll have to use earlycon to
see the crash log.

> The previous patches were supposed to fix the issue but they actually 
> didn't. So how is tested the patch ?

I'm not sure. I believe Matthias test that one on mt6589, maybe it have
different behavior, or different timing.

> Regarding the different fixes for this problem, it sounds like you are 
> proceeding by trial and error.
> 
> Please give a more detailed analysis of the problem and especially check 
> the timer is not altered by the firmware leaving it in a transient state 
> or whatever.

See above for my analysis.
Thanks.

Joe.C