On 10/17/24 08:16, AngeloGioacchino Del Regno wrote:
Il 17/10/24 16:09, Guenter Roeck ha scritto:
On 10/16/24 23:43, yassine.oudjana@xxxxxxxxx wrote:
...
Say I don't want to use the watchdog (which I don't, all I need from TOPRGU is the resets, I don't care about the watchdog). Not starting the watchdog means I can't reset the system because all mtk_wdt_restart will do is make TOPRGU send me an IRQ that I have no use for.
If you don't want to use the watchdog, then you don't need to care about bark
interrupts and you don't need any mtk_wdt_restart() functionality at all :-)
I need mtk_wdt_restart to restart my system. I shouldn't need to take off my phone's back cover and remove the battery every time :)
I think what Guenter said makes sense. We should make sure the watchdog is started when calling mtk_wdt_restart or at least configured in such a way that we are sure it will issue a system reset.
It is more than that. There is no limitation in the watchdog API that says
"you must only use the watchdog kernel driver to reset the system if the
watchdog has been activated from userspace". Such a limitation would be
completely arbitrary and not make any sense. It is perfectly fine to enable
the watchdog from the restart callback if needed. Actually, all restart
handlers in watchdog drivers have to do that if they indeed use a watchdog
to reset the system.
Actually, I am not entirely sure I understand what we are arguing about.
Guenter:
We're arguing about bad configuration and lots of misunderstanding.
Regarding WDT_MODE_EXRST_EN: when enabled, it enables an external output
reset signal - meaning that it's going to flip the state of a GPIO to active
(high in Yassine's case - as that's configured through WDT_MODE BIT(1) and
his 0x5c means that it's flipped on), signaling to another chip (usually,
the PMIC...!) that we want to reset the system.
Explaining what Yassine is doing with this commit: he is flipping the IRQ_EN
bit [BIT(5)] in WDT_MODE.
When bit 5 *is set*, the watchdog bark event will only raise an interrupt and
will not reset the system (that's left to be done to an interrupt handler in
the driver).
When bit 5 *is NOT set*, the watchdog bark event will trigger a CPU reset.
Now, my confusion came from the fact that he's trying to fix a watchdog bark
event so that it triggers system reset, but I didn't understand the actual
reason why he wants to do that - which is powering off the system!
Yassine:
You don't *have to* rely on the watchdog to reset the system, and if you use
only that - especially on a smartphone - I'm mostly sure that you'll get
power leakage.
Before you read the following - please note that this is platform dependent
so, take this with a grain of salt: it is the PMIC that should get configured
to take your system down! I have a hunch that this works for you only because
the platform will reboot, and then the bootloader will decide to turn off the
system for you by default (that, unless you send a warm reboot indication).
That flow looks more like a hack than a solution for an actual problem.
Now - whether you want to fix your platform or not, this is out of the scope
of this commit, which is - in the end - still fixing something that is wrong.
Effectively, as Guenter said, if the watchdog is never started, the restart
function is not going to reboot the system, so yes this problem needs to be
fixed.
There are two problems in this driver that can be solved with:
1. Disable IRQ generation when *no irq* is found in DT; and
2. Implement support for reboot in mtk_wdt_isr() by reading the WDT_STA
register and by then taking appropriate actions.
That won't work because interrupts are likely disabled when the reset callback
executes. The reset handler must not depend on an interrupt. It has to be
self-contained: It has to configure the hardware to issue a reset.
On some systems, that is done by setting the watchdog timeout to a really low value
and enabling it. Others have a special configuration in the watchdog register set
which triggers a reset immediately. If the hardware supports pretimeout, that would
have to be disabled because the idea is to trigger a reset signal, not an interrupt.
To repeat, setting up the system and then waiting for an interrupt to do something
defeats the purpose of the reset handler, which is to issue a reset signal.
Somehow. If that can be done from an interrupt handler, it can and should be done
immediately instead.
Of course my preference would be N.2 because:
- The pretimeout way is already supported in the driver, and if you specify
a pretimeout, then the watchdog will never trigger SYSRST->XRST: this
is actually a bug (IMO!!), as declaring an IRQ in DT means losing reset (!);
- The WDT_STA register tells you more than just "SW/HW watchdog reset asserted"
and that can be extended in the future to support more than that.
However, I recognize that this may be too much work for you, so, if there's no
way for you to properly add support for N.2 - I can chime in.
As for N.1, the solution is simple: check if platform_get_irq_optional fails
and - if it does - force unsetting (WDT_MODE_IRQ_EN | WDT_MODE_DUAL_EN) in
WDT_MODE, if and only if WDT_EN is not set.. but that, in the probe function!
All that should be configured in the reset handler. It has to disable interrupts
even if there is interrupt support because that is not what is wanted at this point.
Thanks,
Guenter
