Message-ID: <CA+zEjCsFif-pT6aW-XwYiSuLsSbBTWVxme6S2rdSN0oQETnZiA@mail.gmail.com>
Date:   Thu, 14 Oct 2021 17:51:22 +0200
From:   Alexandre Ghiti <alexandre.ghiti@...onical.com>
To:     Adam Thomson <Adam.Thomson.Opensource@...semi.com>
Cc:     David Abdurachmanov <david.abdurachmanov@...il.com>,
        Support Opensource <Support.Opensource@...semi.com>,
        Lee Jones <lee.jones@...aro.org>,
        "linux-riscv@...ts.infradead.org" <linux-riscv@...ts.infradead.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] drivers: mfd: da9063: Add restart notifier implementation

Hi Adam,

On Tue, Oct 12, 2021 at 12:33 PM Adam Thomson
<Adam.Thomson.Opensource@...semi.com> wrote:
>
> On 08 October 2021 10:46, Adam Thomson wrote:
>
> > > > Thanks for the info. So we believe, based on the event registers values
> > > > provided, it is the RTC event as that's not cleared by a power-cycle (it's in
> > > > the always-on domain). The other test would be to mask this event
> > > immediately
> > > > after an RTC based reboot and see if the long key-press then shuts down the
> > > > device. I suspect it would in that case, as per you clearing the event.
> > >
> > > Indeed if I mask the RTC alarm in IRQ_MASK_A, the intempestive reboot
> > > disappears. But that's not something we can do: the reset driver will
> > > actually be implemented in openSBI at some point where the devices are
> > > probed on-demand (the same applies to u-boot I think), so we cannot
> > > clear or mask anything at boot.
> > >
> > > >
> > > > I'm still curious as to the 16 seconds though. Would that be when the kernel
> > > > finally starts and masks/clears events (seems quite a long time)?
> > >
> > > No, the kernel is not up yet. Maybe 16sec is not the right value, I
> > > may have been influenced by the discussion here
> > > https://www.dialog-semiconductor.com/products/pmics?post_id=10052#tab-
> > > support_tab_content.
> > >
> > > It seems there is some consensus about having this reset driver be a
> > > SiFive Unmatched board specific thing: quid of the sequence I propose
> > > in this patch then? It does not mess with the RTC registers, it
> > > reboots reliably and there's no intempestive reboot: is it dangerous
> > > to use? Or do you have another alternative?
> >
> > Yes, a board specific implementation would be the way to go. We're just checking
> > through the sequence again to be absolutely sure of any pitfalls that may
> > present themselves, and will get back to you when we have something more.
>
> So having examined your sequence again it's now clearer as to what is happening.
> With the sequence you provided this is only a partial reset whereby all of the
> output rails are sequenced down then up again and restored to OTP voltages.
> However the remainder of the chip settings aren't reset as this isn't a true
> reset of the device going through full reload from OTP, so for example settings
> of regulator mode GPIO states, or IRQ mask bits would persist on the restart,
> which could have implications on system operation.

Ok, it's not perfect but I think those are settings that will get
reinitialized by the corresponding drivers while booting, contrary to
the RTC registers which are clobbered by the other method.

>
> In addition the only bits of interest for you should be:
>
>  - CONTROL_F (0x13)
>         WAKE_UP (BIT 2) = 1
>  - CONTROL_A (0x0E)
>         SYSTEM_EN (BIT 0) = 0
>
> Setting those two bits should be enough to trigger the partial reset sequence.
> The other bits you had in your sequence don't seem to be necessary or relevant.
>
> One final caveat to this approach is that there is a 16s internal timer (as you
> noted before, VDD_START) which is started when the device moves to ACTIVE mode.
> When that 16s timer expires the device will clear the WAKE_UP bit automatically.
> This means there's the outside chance that you could try the reset command
> sequence above around the same time, and that could mean you set the WAKE_UP
> bit, but it's immediately cleared again by this timer expiry before the
> SYSTEM_EN bit is set low. In that case there would be a need for an external
> event (e.g. ONKEY) to kick the system to start again.

Ok, the risk exists but the window is quite small.

After all, the solution I first proposed is not perfect, but now we
know why it works and IMO it has less drawbacks than using the RTC
registers, so I think we should go for this solution. I'll see if I
can help Nikita implement this directly in openSBI.

@Adam Thomson I had migrated the DA9063 device tree bindings to yaml,
I'll push that soon. Thanks for all your help, much appreciated.

Thanks,

Alex

Powered by blists - more mailing lists