Re: [PATCH 1/7] media: hantro: add support for reset lines

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Ezequiel,

W dniu 23.11.2021 o 19:07, Ezequiel Garcia pisze:
Hi all,

Reset logic tends to be highly integration-specific, so it  could be more robust to deal  with this in  the machine specific file. I have some vague recollection of our experience here when we  integrated vc8000 last year, but I cannot recall the outcome.


Do you mean vpu->variant->init()?

That is the very first thing we do after the devm_*() calls. So any subsequent
initialization that fails would want vpu->variant->deinit(). Maybe at this
moment handling the resets at the common level is simpler? Existing drivers
will get NULL anyway from devm_reset_control_array_get().

Regards,

Andrzej

I'm Ccing Bob who might remember better.

Thanks,
Ezequiel



El mar., nov. 23, 2021 1:46 PM, Jernej Škrabec <jernej.skrabec@xxxxxxxxx <mailto:jernej.skrabec@xxxxxxxxx>> escribió:

    Hi all,

    Dne torek, 23. november 2021 ob 17:36:57 CET je Andrzej Pietrasiewicz
    napisal(a):
     > Hi Dan, hi Jernej,
     >
     > W dniu 23.11.2021 o 15:59, Dan Carpenter pisze:
     > > On Tue, Nov 23, 2021 at 12:09:03PM +0100, Andrzej Pietrasiewicz wrote:
     > >>> diff --git a/drivers/staging/media/hantro/hantro_drv.c b/drivers/staging/
    media/hantro/hantro_drv.c
     > >>> index ab2467998d29..8c3de31f51b3 100644
     > >>> --- a/drivers/staging/media/hantro/hantro_drv.c
     > >>> +++ b/drivers/staging/media/hantro/hantro_drv.c
     > >>> @@ -905,6 +905,10 @@ static int hantro_probe(struct platform_device
    *pdev)
     > >>>                           return PTR_ERR(vpu->clocks[0].clk);
     > >>>           }
     > >>> + vpu->resets = devm_reset_control_array_get(&pdev->dev, false,
    true);
     > >>> + if (IS_ERR(vpu->resets))
     > >>> +         return PTR_ERR(vpu->resets);
     > >>> +
     > >>>           num_bases = vpu->variant->num_regs ?: 1;
     > >>>           vpu->reg_bases = devm_kcalloc(&pdev->dev, num_bases,
     > >>>                                         sizeof(*vpu->reg_bases),
    GFP_KERNEL);
     > >>> @@ -978,10 +982,16 @@ static int hantro_probe(struct platform_device
    *pdev)
     > >>>           pm_runtime_use_autosuspend(vpu->dev);
     > >>>           pm_runtime_enable(vpu->dev);
     > >          ^^^^^^^^^^^^^^^^^^^^^^^^^^^
     > > It looks like this is the pm stuff that we have to unwind on error
     > >
     > >>> + ret = reset_control_deassert(vpu->resets);
     > >>> + if (ret) {
     > >>> +         dev_err(&pdev->dev, "Failed to deassert resets\n");
     > >>> +         return ret;
     > >                  ^^^^^^^^^^
     > > So this return should be a goto undo_pm_stuff
     > >
     > >
     > >>> + }
     > >>> +
     > >>>           ret = clk_bulk_prepare(vpu->variant->num_clocks, vpu->clocks);
     > >>>           if (ret) {
     > >>>                   dev_err(&pdev->dev, "Failed to prepare clocks\n");
     > >>> -         return ret;
     > >
     > > And this return should also have been a goto so it's a bug in the
     > > original code.
     >
     > So we probably want a separate patch addressing that first, and then
     > the series proper on top of that.

    I was just about to suggest that.

    Other drivers usually enable PM last, so they don't have PM calls in unwind
    code. However, I think current approach is just as valid (with a fix).

    Best regards,
    Jernej

     >
     > Regards,
     >
     > Andrzej
     >
     > >
     > >>> +         goto err_rst_assert;
     > >>
     > >> Before your patch is applied if clk_bulk_prepare() fails, we
     > >> simply return on the spot. After the patch is applied not only
     > >> do you...
     > >>
     > >>>           }
     > >>>           ret = v4l2_device_register(&pdev->dev, &vpu->v4l2_dev);
     > >>> @@ -1037,6 +1047,8 @@ static int hantro_probe(struct platform_device
    *pdev)
     > >>>           v4l2_device_unregister(&vpu->v4l2_dev);
     > >>>    err_clk_unprepare:
     > >>>           clk_bulk_unprepare(vpu->variant->num_clocks, vpu->clocks);
     > >>> +err_rst_assert:
     > >>> + reset_control_assert(vpu->resets);
     > >>
     > >> ...revert the effect of reset_control_deassert(), you also...
     > >>
     > >>>           pm_runtime_dont_use_autosuspend(vpu->dev);
     > >>>           pm_runtime_disable(vpu->dev);
     > >>
     > >> ... do pm_*() stuff. Is there any reason why this is needed?
     > >
     > > So, yes, it's needed, but you're correct to spot that it's not
     > > consistent.
     > >
     > > regards,
     > > dan carpenter
     > >
     >
     >







[Index of Archives]     [Linux Driver Development]     [Linux Driver Backports]     [DMA Engine]     [Linux GPIO]     [Linux SPI]     [Video for Linux]     [Linux USB Devel]     [Linux Coverity]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Yosemite Backpacking]
  Powered by Linux