2018-12-19 19:12:29

by Matthias Kaehlcke

[permalink] [raw]
Subject: [PATCH] tty: serial: qcom_geni_serial: Fix wrap around of TX buffer

Before commit a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix
softlock") the size of TX transfers was limited to the TX FIFO size,
and wrap arounds of the UART circular buffer were split into two
transfers. With the commit wrap around are allowed within a transfer.
The TX FIFO of the geni serial port uses a word size of 4 bytes. In
case of a circular buffer wrap within a transfer the driver currently
may write an incomplete word to the FIFO, with some bytes containing
data from the circular buffer and others being zero. Since the
transfer isn't completed yet the zero bytes are sent as if they were
actual data.

Handle wrap arounds of the TX buffer properly and ensure that words
written to the TX FIFO always contain valid data (unless the transfer
is completed).

Fixes: a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix softlock")
Signed-off-by: Matthias Kaehlcke <[email protected]>
---
drivers/tty/serial/qcom_geni_serial.c | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/drivers/tty/serial/qcom_geni_serial.c b/drivers/tty/serial/qcom_geni_serial.c
index cf7a95e339ad9..2ee2d3286a6b4 100644
--- a/drivers/tty/serial/qcom_geni_serial.c
+++ b/drivers/tty/serial/qcom_geni_serial.c
@@ -744,7 +744,7 @@ static void qcom_geni_serial_handle_tx(struct uart_port *uport, bool done,
avail *= port->tx_bytes_pw;

tail = xmit->tail;
- chunk = min3(avail, pending, (size_t)(UART_XMIT_SIZE - tail));
+ chunk = min(avail, pending);
if (!chunk)
goto out_write_wakeup;

@@ -766,19 +766,21 @@ static void qcom_geni_serial_handle_tx(struct uart_port *uport, bool done,

memset(buf, 0, ARRAY_SIZE(buf));
tx_bytes = min_t(size_t, remaining, port->tx_bytes_pw);
- for (c = 0; c < tx_bytes ; c++)
- buf[c] = xmit->buf[tail + c];
+
+ for (c = 0; c < tx_bytes ; c++) {
+ buf[c] = xmit->buf[tail++];
+ tail &= UART_XMIT_SIZE - 1;
+ }

iowrite32_rep(uport->membase + SE_GENI_TX_FIFOn, buf, 1);

i += tx_bytes;
- tail += tx_bytes;
uport->icount.tx += tx_bytes;
remaining -= tx_bytes;
port->tx_remaining -= tx_bytes;
}

- xmit->tail = tail & (UART_XMIT_SIZE - 1);
+ xmit->tail = tail;

/*
* The tx fifo watermark is level triggered and latched. Though we had
--
2.20.0.405.gbc1bbc6f85-goog



2018-12-19 19:20:55

by Evan Green

[permalink] [raw]
Subject: Re: [PATCH] tty: serial: qcom_geni_serial: Fix wrap around of TX buffer

On Wed, Dec 19, 2018 at 10:17 AM Matthias Kaehlcke <[email protected]> wrote:
>
> Before commit a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix
> softlock") the size of TX transfers was limited to the TX FIFO size,
> and wrap arounds of the UART circular buffer were split into two
> transfers. With the commit wrap around are allowed within a transfer.
> The TX FIFO of the geni serial port uses a word size of 4 bytes. In
> case of a circular buffer wrap within a transfer the driver currently
> may write an incomplete word to the FIFO, with some bytes containing
> data from the circular buffer and others being zero. Since the
> transfer isn't completed yet the zero bytes are sent as if they were
> actual data.
>
> Handle wrap arounds of the TX buffer properly and ensure that words
> written to the TX FIFO always contain valid data (unless the transfer
> is completed).
>
> Fixes: a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix softlock")
> Signed-off-by: Matthias Kaehlcke <[email protected]>

Oh nice, so this did end up being your problem with corrupt
characters? Strange though, I still would have expected this bug to
result in inserted serial characters, rather than edited ones.
Either way, this looks good to me.

Reviewed-by: Evan Green <[email protected]>

2018-12-19 19:21:03

by Ryan Case

[permalink] [raw]
Subject: Re: [PATCH] tty: serial: qcom_geni_serial: Fix wrap around of TX buffer

On Wed, Dec 19, 2018 at 10:50 AM Evan Green <[email protected]> wrote:
>
> On Wed, Dec 19, 2018 at 10:17 AM Matthias Kaehlcke <[email protected]> wrote:
> >
> > Before commit a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix
> > softlock") the size of TX transfers was limited to the TX FIFO size,
> > and wrap arounds of the UART circular buffer were split into two
> > transfers. With the commit wrap around are allowed within a transfer.
> > The TX FIFO of the geni serial port uses a word size of 4 bytes. In
> > case of a circular buffer wrap within a transfer the driver currently
> > may write an incomplete word to the FIFO, with some bytes containing
> > data from the circular buffer and others being zero. Since the
> > transfer isn't completed yet the zero bytes are sent as if they were
> > actual data.
> >
> > Handle wrap arounds of the TX buffer properly and ensure that words
> > written to the TX FIFO always contain valid data (unless the transfer
> > is completed).
> >
> > Fixes: a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix softlock")
> > Signed-off-by: Matthias Kaehlcke <[email protected]>
>
> Oh nice, so this did end up being your problem with corrupt
> characters? Strange though, I still would have expected this bug to
> result in inserted serial characters, rather than edited ones.
> Either way, this looks good to me.
>
> Reviewed-by: Evan Green <[email protected]>

UART bluetooth firmware downloads consistently working for me again.

Tested-by: Ryan Case <[email protected]>

2018-12-19 19:21:28

by Matthias Kaehlcke

[permalink] [raw]
Subject: Re: [PATCH] tty: serial: qcom_geni_serial: Fix wrap around of TX buffer

On Wed, Dec 19, 2018 at 10:49:26AM -0800, Evan Green wrote:
> On Wed, Dec 19, 2018 at 10:17 AM Matthias Kaehlcke <[email protected]> wrote:
> >
> > Before commit a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix
> > softlock") the size of TX transfers was limited to the TX FIFO size,
> > and wrap arounds of the UART circular buffer were split into two
> > transfers. With the commit wrap around are allowed within a transfer.
> > The TX FIFO of the geni serial port uses a word size of 4 bytes. In
> > case of a circular buffer wrap within a transfer the driver currently
> > may write an incomplete word to the FIFO, with some bytes containing
> > data from the circular buffer and others being zero. Since the
> > transfer isn't completed yet the zero bytes are sent as if they were
> > actual data.
> >
> > Handle wrap arounds of the TX buffer properly and ensure that words
> > written to the TX FIFO always contain valid data (unless the transfer
> > is completed).
> >
> > Fixes: a1fee899e5bed ("tty: serial: qcom_geni_serial: Fix softlock")
> > Signed-off-by: Matthias Kaehlcke <[email protected]>
>
> Oh nice, so this did end up being your problem with corrupt
> characters?

Yes

> Strange though, I still would have expected this bug to
> result in inserted serial characters, rather than edited ones.

You are right here, initially I interpreted that some bytes changed,
but actually we observed inserted bytes, and later missing ones at the
end of the transfer.

> Either way, this looks good to me.
>
> Reviewed-by: Evan Green <[email protected]>

Thanks!