* Chen-Yu Tsai <wenst@xxxxxxxxxxxx> [230605 13:01]:
> On Mon, Jun 5, 2023 at 8:24 PM Tony Lindgren <tony@xxxxxxxxxxx> wrote:
> >
> > * Chen-Yu Tsai <wenst@xxxxxxxxxxxx> [230605 11:34]:
> > > On Mon, Jun 5, 2023 at 2:15 PM Tony Lindgren <tony@xxxxxxxxxxx> wrote:
> > > > diff --git a/drivers/tty/serial/8250/8250_mtk.c b/drivers/tty/serial/8250/8250_mtk.c
> > > > --- a/drivers/tty/serial/8250/8250_mtk.c
> > > > +++ b/drivers/tty/serial/8250/8250_mtk.c
> > > > @@ -425,11 +439,10 @@ mtk8250_set_termios(struct uart_port *port, struct ktermios *termios,
> > > > static int __maybe_unused mtk8250_runtime_suspend(struct device *dev)
> > > > {
> > > > struct mtk8250_data *data = dev_get_drvdata(dev);
> > > > - struct uart_8250_port *up = serial8250_get_port(data->line);
> > > >
> > > > /* wait until UART in idle status */
> > > > while
> > > > - (serial_in(up, MTK_UART_DEBUG0));
> > > > + (mtk8250_read(data, MTK_UART_DEBUG0));
> > >
> > > I believe it still gets stuck here sometimes.
> >
> > Hmm so maybe you need to mtk8250_write(data, 0, MTK_UART_RATE_FIX) in
> > probe before pm_runtime_resume_and_get() that enables the baud clock?
> > That's something I changed, so maybe it messes up things.
>
> I think it has something to do with the do_pm() function calling
> the callbacks directly, then also calling runtime PM.
Yeah I'm not following really what's going on there.. So then I guess the
call for mtk8250_write(data, 0, MTK_UART_RATE_FIX) should be after the
pm_runtime_resume_and_get() call.
> > Looking at the 8250_mtk git log, it's runtime PM functions seem to only
> > currently manage the baud clock so register access should be doable
> > without runtime PM resume?
>
> Actually it only manages the bus clock. The baud clock is simply the system
> XTAL which is not gateble.
OK
> > > With your earlier patch, it could get through registering the port, and
> > > the console would show
> > >
> > > 11002000.serial: ttyS0 at MMIO 0x11002000 (irq = 240, base_baud =
> > > 1625000) is a ST16650V2
> > >
> > > for the console UART.
> >
> > OK
> >
> > > Angelo mentioned that we should be using SLEEP_REQ/SLEEP_ACK registers
> > > in the MTK UART hardware.
> > >
> > > I tried reworking it into your patch here, but it causes issues with the
> > > UART-based Bluetooth on one of my devices. After the UART runtime suspends
> > > and resumes, something is off and causes the transfers during Bluetooth
> > > init to become corrupt.
> > >
> > > I'll try some more stuff, but the existing code seems timing dependent.
> > > If I add too many printk statements to the runtime suspend/resume
> > > callbacks, things seem to work. One time I even ended up with broken
> > > UARTs but otherwise booted up the system.
> >
> > Well another thing that now changes is that we now runtime suspend the
> > port at the end of the probe. What the 8250_mtk probe was doing earlier
> > it was leaving the port baud clock enabled, but runtime PM disabled
> > until mtk8250_do_pm() I guess.
>
> I guess that's the biggest difference? Since the *bus* clock gets disabled,
> any access will hang. Is it enough to just support runtime PM? Or do I have
> to also have UART_CAP_RPM?
Maybe try changing pm_runtime_put_sync() at the end of the probe to just
pm_runtime_put_noidle()? Then the driver should be back to where it was
with clocks enabled but runtime PM suspended.
I don't think you need UART_CAP_RPM right now unless 8250_mtk adds support
for autosuspend. That stuff will get replaced by the serial_core generic
PM patch from Andy. I think in it's current form 8250_mtk just gets enabled
when the port is opened, and disabled when the port is closed. And gets
disabled for system suspend.
Regards,
Tony
