Re: [PATCH] tty: serial: qcom_geni_serial: Fix softlock

[Ryan Case <ryandcase@chromium.org>]

On Tue, Nov 27, 2018 at 4:20 PM Stephen Boyd <swboyd@chromium.org> wrote:
>
> Quoting Ryan Case (2018-11-26 18:25:36)
> > Transfers were being divided into device FIFO sized (64 byte max)
> > operations which would poll for completion within a spin_lock_irqsave /
> > spin_unlock_irqrestore block. This both made things slow by waiting for
> > the FIFO to completely drain before adding further data and would also
> > result in softlocks on large transmissions.
> >
> > This patch allows larger transfers with continuous FIFO additions as
> > space becomes available and removes polling from the interrupt handler.
> >
> > Signed-off-by: Ryan Case <ryandcase@chromium.org>
> > Version: 1
>
> I've never seen a Version tag before. Did you manually add this?

I submitted with patman, this should have been Series-version:

>
> > diff --git a/drivers/tty/serial/qcom_geni_serial.c b/drivers/tty/serial/qcom_geni_serial.c
> > index 7ded51081add..835a184e0b7d 100644
> > --- a/drivers/tty/serial/qcom_geni_serial.c
> > +++ b/drivers/tty/serial/qcom_geni_serial.c
> > @@ -117,6 +117,8 @@ struct qcom_geni_serial_port {
> >         u32 *rx_fifo;
> >         u32 loopback;
> >         bool brk;
> > +
> > +       u32 cur_tx_remaining;
>
> Nitpick: Can it just be tx_remaining? And why u32? Why not unsigned int?

Sure, and unsigned int is fine.

>
> >  };
> >
> >  static const struct uart_ops qcom_geni_console_pops;
> > @@ -439,6 +441,7 @@ static void qcom_geni_serial_console_write(struct console *co, const char *s,
> >         struct qcom_geni_serial_port *port;
> >         bool locked = true;
> >         unsigned long flags;
> > +       unsigned int geni_status;
>
> Nitpick: Use u32 for register reads.

will do.

>
> >
> >         WARN_ON(co->index < 0 || co->index >= GENI_UART_CONS_PORTS);
> >
> > @@ -465,9 +470,17 @@ static void qcom_geni_serial_console_write(struct console *co, const char *s,
> >                 }
> >                 writel_relaxed(M_CMD_CANCEL_EN, uport->membase +
> >                                                         SE_GENI_M_IRQ_CLEAR);
> > -       }
> > +       } else if ((geni_status & M_GENI_CMD_ACTIVE) && !port->cur_tx_remaining)
> > +               /* It seems we can interrupt existing transfers unless all data
>
> Nitpick: Have /* on a line by itself
>
> Is this comment supposed to say "we can't interrupt existing transfers"?

Nope, comment is correct as is.

>
> > +                * has been sent, in which case we need to look for done first.
> > +                */
> > +               qcom_geni_serial_poll_tx_done(uport);
>
> Another nitpick: Please put braces around multi-line if branches for
> greater code clarity.

will do.

>
> >
> >         __qcom_geni_serial_console_write(uport, s, count);
> > +
> > +       if (port->cur_tx_remaining)
> > +               qcom_geni_serial_setup_tx(uport, port->cur_tx_remaining);
>
> Does this happen? Is the console being used as a tty at the same time?

Yup, happens quite a bit.

>
> > +
> >         if (locked)
> >                 spin_unlock_irqrestore(&uport->lock, flags);
> >  }
> > @@ -701,40 +714,47 @@ static void qcom_geni_serial_handle_rx(struct uart_port *uport, bool drop)
> >         port->handle_rx(uport, total_bytes, drop);
> >  }
> >
> > -static void qcom_geni_serial_handle_tx(struct uart_port *uport)
> > +static void qcom_geni_serial_handle_tx(struct uart_port *uport, bool done,
> > +               bool active)
> >  {
> >         struct qcom_geni_serial_port *port = to_dev_port(uport, uport);
> >         struct circ_buf *xmit = &uport->state->xmit;
> >         size_t avail;
> >         size_t remaining;
> > +       size_t pending;
> >         int i;
> >         u32 status;
> >         unsigned int chunk;
> >         int tail;
> > -       u32 irq_en;
> >
> > -       chunk = uart_circ_chars_pending(xmit);
> >         status = readl_relaxed(uport->membase + SE_GENI_TX_FIFO_STATUS);
> > -       /* Both FIFO and framework buffer are drained */
> > -       if (!chunk && !status) {
> > +
> > +       /* Complete the current tx command before taking newly added data */
> > +       if (active)
> > +               pending = port->cur_tx_remaining;
> > +       else
> > +               pending = uart_circ_chars_pending(xmit);
> > +
> > +       /* All data has been transmitted and acknowledged as received */
> > +       if (!pending && !status && done) {
>
> Nitpick: status is a poor variable name to test here. I don't understand
> what this line is doing. Maybe it would help to have another local
> variable like 'needs_attention'?

It could be renamed but since this isn't a general file cleanup patch
I was avoiding non-functional changes. It is the TX_FIFO_STATUS
register, if non-zero there is still data in the FIFO or related
activity ongoing.

>
> >                 qcom_geni_serial_stop_tx(uport);
> >                 goto out_write_wakeup;
> >         }
> >
> > -       if (!uart_console(uport)) {
> > -               irq_en = readl_relaxed(uport->membase + SE_GENI_M_IRQ_EN);
> > -               irq_en &= ~(M_TX_FIFO_WATERMARK_EN);
> > -               writel_relaxed(0, uport->membase + SE_GENI_TX_WATERMARK_REG);
> > -               writel_relaxed(irq_en, uport->membase + SE_GENI_M_IRQ_EN);
> > -       }
> > +       avail = port->tx_fifo_depth - (status & TX_FIFO_WC);
> > +       avail *= port->tx_bytes_pw;
> > +       if (avail < 0)
> > +               avail = 0;
>
> How can 'avail' be less than 0? It's size_t which is unsigned? If
> underflow is happening from that subtraction or overflow from the
> multiply that could be bad but I hope that is impossible.

I hope underflow is impossible as well. However, if the hardware did
wind up in a strange state I wanted to err on the side of not throwing
away data and being able to resume later if things recovered. I can
remove the defensive checks if that's the custom, otherwise I'll
update the comparison logic accordingly.

>
> >
> > -       avail = (port->tx_fifo_depth - port->tx_wm) * port->tx_bytes_pw;
> >         tail = xmit->tail;
> > -       chunk = min3((size_t)chunk, (size_t)(UART_XMIT_SIZE - tail), avail);
> > +       chunk = min3((size_t)pending, (size_t)(UART_XMIT_SIZE - tail), avail);
>
> Nitpick: If we made 'avail' unsigned int would we be able to drop the
> casts on this min3() call? This line is quite hard to read.

Seems they can go away without any changes.

>
> >         if (!chunk)
> >                 goto out_write_wakeup;
> >
> > -       qcom_geni_serial_setup_tx(uport, chunk);
> > +       if (!port->cur_tx_remaining) {
> > +               qcom_geni_serial_setup_tx(uport, pending);
> > +               port->cur_tx_remaining = pending;
> > +       }
> >
> >         remaining = chunk;
> >         for (i = 0; i < chunk; ) {
> > @@ -767,6 +786,7 @@ static irqreturn_t qcom_geni_serial_isr(int isr, void *dev)
> >  {
> >         unsigned int m_irq_status;
> >         unsigned int s_irq_status;
> > +       unsigned int geni_status;
>
> Nitpick: I guess this driver isn't using u32 for registers already.
> Would be nice to mop this up in another patch.
>
> >         struct uart_port *uport = dev;
> >         unsigned long flags;
> >         unsigned int m_irq_en;
> >