Re: [PATCH v4 3/3] dmaengine: sf-pdma: Get number of channel by device tree

[Zong Li]

On Fri, Jan 21, 2022 at 4:33 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
>
> Hi Zong, Palmer,
>
> On Fri, Jan 21, 2022 at 3:21 AM Zong Li <zong.li@xxxxxxxxxx> wrote:
> > On Fri, Jan 21, 2022 at 2:52 AM Palmer Dabbelt <palmer@xxxxxxxxxxx> wrote:
> > > On Sun, 16 Jan 2022 17:35:28 PST (-0800), zong.li@xxxxxxxxxx wrote:
> > > > It currently assumes that there are always four channels, it would
> > > > cause the error if there is actually less than four channels. Change
> > > > that by getting number of channel from device tree.
> > > >
> > > > For backwards-compatible, it uses the default value (i.e. 4) when there
> > > > is no 'dma-channels' information in dts.
> > >
> > > Some of the same wording issues here as those I pointed out in the DT
> > > bindings patch.
> > >
> > > > Signed-off-by: Zong Li <zong.li@xxxxxxxxxx>
>
> > > > --- a/drivers/dma/sf-pdma/sf-pdma.c
> > > > +++ b/drivers/dma/sf-pdma/sf-pdma.c
> > > > @@ -482,9 +482,7 @@ static void sf_pdma_setup_chans(struct sf_pdma *pdma)
> > > >  static int sf_pdma_probe(struct platform_device *pdev)
> > > >  {
> > > >       struct sf_pdma *pdma;
> > > > -     struct sf_pdma_chan *chan;
> > > >       struct resource *res;
> > > > -     int len, chans;
> > > >       int ret;
> > > >       const enum dma_slave_buswidth widths =
> > > >               DMA_SLAVE_BUSWIDTH_1_BYTE | DMA_SLAVE_BUSWIDTH_2_BYTES |
> > > > @@ -492,13 +490,21 @@ static int sf_pdma_probe(struct platform_device *pdev)
> > > >               DMA_SLAVE_BUSWIDTH_16_BYTES | DMA_SLAVE_BUSWIDTH_32_BYTES |
> > > >               DMA_SLAVE_BUSWIDTH_64_BYTES;
> > > >
> > > > -     chans = PDMA_NR_CH;
> > > > -     len = sizeof(*pdma) + sizeof(*chan) * chans;
> > > > -     pdma = devm_kzalloc(&pdev->dev, len, GFP_KERNEL);
> > > > +     pdma = devm_kzalloc(&pdev->dev, sizeof(*pdma), GFP_KERNEL);
> > > >       if (!pdma)
> > > >               return -ENOMEM;
> > > >
> > > > -     pdma->n_chans = chans;
> > > > +     ret = of_property_read_u32(pdev->dev.of_node, "dma-channels",
> > > > +                                &pdma->n_chans);
> > > > +     if (ret) {
> > > > +             dev_notice(&pdev->dev, "set number of channels to default value: 4\n");
> > > > +             pdma->n_chans = PDMA_MAX_NR_CH;
> > > > +     }
> > > > +
> > > > +     if (pdma->n_chans > PDMA_MAX_NR_CH) {
> > > > +             dev_err(&pdev->dev, "the number of channels exceeds the maximum\n");
> > > > +             return -EINVAL;
> > >
> > > Can we get away with just using only the number of channels the driver
> > > actually supports?  ie, just never sending an op to the channels above
> > > MAX_NR_CH?  That should leave us with nothing to track.
>
> In theory we can...
>
> > It might be a bit like when pdma->n_chans is bigger than the maximum,
> > set the pdma->chans to PDMA_MAX_NR_CH, then we could ensure that we
> > don't access the channels above the maximum. If I understand
> > correctly, I gave the similar thought in the thread of v2 patch, and
> > there are some discussions on that, but this way seems to lead to
> > hard-to-track problems.
>
> ... but that would mean that when a new variant appears that supports
> more channels, no error is printed, and people might not notice
> immediately that the higher channels are never used.
>

I guess people might need to follow the dt-bindings, so they couldn't
specify the number of channels to the value which is more than
maximum. But as you mentioned, if people don't notice that and specify
it more than maximum,  they wouldn't be aware that the higher channels
are never used. It seems to me that we could keep returning the error
there, or show a warning message and use PDMA_MAX_NR_CH in that
situation, both looks good to me.

> 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