Hi Kaneko-san, Matsuoka-san, Mizuguchi-san,
On Fri, May 1, 2015 at 7:11 PM, Yoshihiro Kaneko <ykaneko0929@xxxxxxxxx> wrote:
> From: Koji Matsuoka <koji.matsuoka.xm@xxxxxxxxxxx>
>
> Signed-off-by: Koji Matsuoka <koji.matsuoka.xm@xxxxxxxxxxx>
> Signed-off-by: Kazuya Mizuguchi <kazuya.mizuguchi.ks@xxxxxxxxxxx>
> Signed-off-by: Yoshihiro Kaneko <ykaneko0929@xxxxxxxxx>
Thanks for your patch!
BTW, with which DMA engine was this tested? The old shdmac or the new
rcar-dmac?
> --- a/drivers/tty/serial/sh-sci.c
> +++ b/drivers/tty/serial/sh-sci.c
> @@ -1348,13 +1363,26 @@ static void sci_rx_dma_release(struct sci_port *s, bool enable_pio)
> {
> struct dma_chan *chan = s->chan_rx;
> struct uart_port *port = &s->port;
> + size_t buf_len_rx;
> + int t = 100000;
> +
> + s->rx_release_flag = 1;
> + while (t--) {
> + if (!s->rx_flag)
> + break;
> + usleep_range(10, 50);
> + }
This is done to make sure rx DMA has completed before releasing the channel
(and setting s->chan_rx = NULL)?
Is there a better way to wait for this, without looping?
> s->chan_rx = NULL;
> @@ -1419,11 +1457,19 @@ static void work_fn_rx(struct work_struct *work)
> struct uart_port *port = &s->port;
> struct dma_async_tx_descriptor *desc;
> int new;
> + int next;
> +
> + if (s->chan_rx == NULL) {
> + dev_dbg(port->dev, "%s: DMA channel is released.\n", __func__);
> + return;
> + }
I believe this is a workaround for the race condition where work_fn_rx()
is called after shutdown of the port, and s->chan_rx has already been set
to NULL in sci_rx_dma_release(), which leads to a NULL pointer dereference
in work_fn_rx()?
As this is a bug fix, it should be a separate patch.
Can this still happen in the presence of the loop until !rx_flag in
sci_rx_dma_release() above?
> @@ -1707,14 +1770,22 @@ static void sci_request_dma(struct uart_port *port)
> chan = dma_request_channel(mask, filter, param);
> dev_dbg(port->dev, "%s: RX: got channel %p\n", __func__, chan);
> if (chan) {
> - dma_addr_t dma[2];
> - void *buf[2];
> + dma_addr_t dma[3];
> + void *buf[3];
> + size_t sum_buf_len;
> + int nr_sg = 2;
> int i;
>
> s->chan_rx = chan;
>
> s->buf_len_rx = 2 * max(16, (int)port->fifosize);
> - buf[0] = dma_alloc_coherent(port->dev, s->buf_len_rx * 2,
> + if (port->type == PORT_SCIF || port->type == PORT_HSCIF) {
> + nr_sg = 3;
> + sum_buf_len = 1 + s->buf_len_rx * 2;
> + } else {
> + sum_buf_len = s->buf_len_rx * 2;
> + }
> + buf[0] = dma_alloc_coherent(port->dev, sum_buf_len,
> &dma[0], GFP_KERNEL);
I think the code manipulating the buffers can be simplified a lot by
storing in sci_port:
- The number of buffers in use (2 vs. 3),
- The sizes of the individual buffers.
That way you can remove many checks for (H)SCIF, and avoid recalculating the
buffer sizes everywhere.
> @@ -1778,7 +1863,6 @@ static int sci_startup(struct uart_port *port)
>
> spin_lock_irqsave(&port->lock, flags);
> sci_start_tx(port);
> - sci_start_rx(port);
Why is this needed?
> @@ -1796,6 +1880,8 @@ static void sci_shutdown(struct uart_port *port)
> sci_stop_tx(port);
> spin_unlock_irqrestore(&port->lock, flags);
>
> + serial_port_out(port, SCSCR, 0x00);
Isn't this already handled by sci_stop_tx() above, and by sci_stop_rx() in
the line above that (not visible in the patch)?
Which additional bits do you need to clear?
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
--
To unsubscribe from this list: send the line "unsubscribe linux-serial" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
