Re: [PATCH 1/8] serial: qcom-geni: fix fifo polling timeout

[Johan Hovold <johan@kernel.org>]

On Wed, Sep 04, 2024 at 02:50:57PM -0700, Doug Anderson wrote:
> On Mon, Sep 2, 2024 at 8:26 AM Johan Hovold <johan+linaro@kernel.org> wrote:
> >
> > The qcom_geni_serial_poll_bit() can be used to wait for events like
> > command completion and is supposed to wait for the time it takes to
> > clear a full fifo before timing out.
> >
> > As noted by Doug, the current implementation does not account for start,
> > stop and parity bits when determining the timeout. The helper also does
> > not currently account for the shift register and the two-word
> > intermediate transfer register.
> >
> > Instead of determining the fifo timeout on every call, store the timeout
> > when updating it in set_termios() and wait for up to 19/16 the time it
> > takes to clear the 16 word fifo to account for the shift and
> > intermediate registers. Note that serial core has already added a 20 ms
> > margin to the fifo timeout.
> >
> > Also note that the current uart_fifo_timeout() interface does
> > unnecessary calculations on every call and also did not exists in
> > earlier kernels so only store its result once. This also facilitates
> > backports as earlier kernels can derive the timeout from uport->timeout,
> > which has since been removed.

> > @@ -270,22 +270,21 @@ static bool qcom_geni_serial_poll_bit(struct uart_port *uport,
> >  {
> >         u32 reg;
> >         struct qcom_geni_serial_port *port;
> > -       unsigned int baud;
> > -       unsigned int fifo_bits;
> >         unsigned long timeout_us = 20000;
> >         struct qcom_geni_private_data *private_data = uport->private_data;
> >
> >         if (private_data->drv) {
> >                 port = to_dev_port(uport);
> > -               baud = port->baud;
> > -               if (!baud)
> > -                       baud = 115200;
> > -               fifo_bits = port->tx_fifo_depth * port->tx_fifo_width;
> > +
> >                 /*
> > -                * Total polling iterations based on FIFO worth of bytes to be
> > -                * sent at current baud. Add a little fluff to the wait.
> > +                * Wait up to 19/16 the time it would take to clear a full
> > +                * FIFO, which accounts for the three words in the shift and
> > +                * intermediate registers.
> > +                *
> > +                * Note that fifo_timeout_us already has a 20 ms margin.
> >                  */
> > -               timeout_us = ((fifo_bits * USEC_PER_SEC) / baud) + 500;
> > +               if (port->fifo_timeout_us)
> > +                       timeout_us = 19 * port->fifo_timeout_us / 16;
> 
> It made me giggle a bit that part of the justification for caching
> "fifo_timeout_us" was to avoid calculations each time through the
> function. ...but then the code does the "19/16" math here instead of
> just including it in the cache. ;-) ;-) ;-)

Heh, yeah, but I was really talking about uart_fifo_timeout() doing
unnecessary calculations on each call (and that value used to be
calculated once and stored for later use).

I also realised that we need to account for the intermediate register
after I wrote the initial commit message, and before that this was just
a shift and add.

> That being said, I'm not really a fan of the "19 / 16" anyway. The 16
> value is calculated elsewhere in the code as:
> 
> port->tx_fifo_depth = geni_se_get_tx_fifo_depth(&port->se);
> port->tx_fifo_width = geni_se_get_tx_fifo_width(&port->se);
> port->rx_fifo_depth = geni_se_get_rx_fifo_depth(&port->se);
> uport->fifosize =
>   (port->tx_fifo_depth * port->tx_fifo_width) / BITS_PER_BYTE;
> 
> ...and here you're just hardcoding it to 16. Then there's also the
> fact that the "19 / 16" will also multiply the 20 ms "slop" added by
> uart_fifo_timeout() which doesn't seem ideal.

Indeed, and the early console code also hardcodes this to 16.

I don't care about the slop being 20 ms or 23.5, this is just a timeout
for the error case.

This will over count a bit if there is uart hw with 256 B fifos, but
could potentially undercount if there is hw with less than 16 words. I'm
not sure if such hw exists, but I'll see what I can find out.

> How about this: we just change "uport->fifosize" to account for the 3
> extra words? So it can be:
> 
> ((port->tx_fifo_depth + 3) * port->tx_fifo_width) / BITS_PER_BYTE;
> 
> ...then the cache will be correct and everything will work out. What
> do you think?

I don't think uart_fifo_timeout traditionally accounts for the shift
register and we wait up to *twice* the time it takes to clear to fifo
anyway (in wait_until_sent). The intermediate register I found here
could perhaps be considered part of the fifo however.

I'll give this some more thought.

Johan