On Mon, 25 Jan 2021 12:21:35 +0100
Ahmad Fatoum <a.fatoum@xxxxxxxxxxxxxx> wrote:
> Hello Jonathan,
>
> On 24.01.21 16:22, Jonathan Cameron wrote:
> > On Fri, 22 Jan 2021 12:33:55 +0100
> > Ahmad Fatoum <a.fatoum@xxxxxxxxxxxxxx> wrote:
> >
> >> For non-DMA usage, we have an easy way to associate a timestamp with a
> >> sample: iio_pollfunc_store_time stores a timestamp in the primary
> >> trigger IRQ handler and stm32_adc_trigger_handler runs in the IRQ thread
> >> to push out the buffer along with the timestamp.
> >>
> >> For this to work, the driver needs to register an IIO_TIMESTAMP channel.
> >> Do this.
> >>
> >> For DMA, it's not as easy, because we don't push the buffers out of
> >> stm32_adc_trigger, but out of stm32_adc_dma_buffer_done, which runs in
> >> a tasklet scheduled after a DMA completion.
> >>
> >> Preferably, the DMA controller would copy us the timestamp into that buffer
> >> as well. Until this is implemented, restrict timestamping support to
> >> only PIO. For low-frequency sampling, PIO is probably good enough.
> >>
> >> Cc: Holger Assmann <has@xxxxxxxxxxxxxx>
> >> Acked-by: Fabrice Gasnier <fabrice.gasnier@xxxxxxxxxxx>
> >> Signed-off-by: Ahmad Fatoum <a.fatoum@xxxxxxxxxxxxxx>
> >
> > This patch itself is fine, but it will expose a potential bug.
> >
> > The buffer passed to iio_push_to_buffers_with_timestamp needs to be suitably
> > aligned to take an 8 byte timestamp and large enough to do so.
> > Currently, in this driver it isn't.
> > u16 buffer[STM32_ADC_MAX_SQ];
> > Appears to be the same length as the channel count, and isn't 8 byte
> > aligned. (add __aligned(8) to fix that)
> >
> > Could you add that fix to this patch as well?
>
> Just done so, thanks. But I think it's very surprising API to expect a void *
> to have a specific alignment. Should this perhaps be encoded into
> the function signature? e.g.
>
> typedef void __aligned_u64_void __aligned(8);
> static inline int iio_push_to_buffers_with_timestamp(struct iio_dev *indio_dev,
> __aligned_u64_void *data, int64_t timestamp);
>
Hmm. Didn't realise you could put __aligned() on a parameter. May well make sense
but we can't do it yet. There are a few left over drivers that don't yet
guarantee the alignment. On some architectures that's fine. Once we've
done a final pass and fixed those up, we can tidy this up.
Note that we had this problem for a good 5-10 years before one report of actually
running into a problem.
> [I assume put_unaligned_* isn't used for performance reasons?)
No, it's actually about expectations of where that buffer might go. There
are lots of potential consumers and many of them can assume a fixed
structure and hence assume alignment.
I agree, it's a very odd ABI requirement :(
Would probably have done things differently if we'd registered the built in
alignment issue earlier.
Jonathan
>
> Cheers,
> Ahmad
>
> >
> > Thanks,
> >
> > Jonathan
> >
> >
> >> ---
> >> v1 -> v2:
> >> - Added comment about timestamping being PIO only (Fabrice)
> >> - Added missing DMA resource clean up in error path (Fabrice)
> >> - Added Fabrice's Acked-by
> >> ---
> >> drivers/iio/adc/stm32-adc.c | 35 +++++++++++++++++++++++++++++------
> >> 1 file changed, 29 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/drivers/iio/adc/stm32-adc.c b/drivers/iio/adc/stm32-adc.c
> >> index c067c994dae2..885bb514503c 100644
> >> --- a/drivers/iio/adc/stm32-adc.c
> >> +++ b/drivers/iio/adc/stm32-adc.c
> >> @@ -1718,7 +1718,7 @@ static void stm32_adc_chan_init_one(struct iio_dev *indio_dev,
> >> }
> >> }
> >>
> >> -static int stm32_adc_chan_of_init(struct iio_dev *indio_dev)
> >> +static int stm32_adc_chan_of_init(struct iio_dev *indio_dev, bool timestamping)
> >> {
> >> struct device_node *node = indio_dev->dev.of_node;
> >> struct stm32_adc *adc = iio_priv(indio_dev);
> >> @@ -1766,6 +1766,9 @@ static int stm32_adc_chan_of_init(struct iio_dev *indio_dev)
> >> return -EINVAL;
> >> }
> >>
> >> + if (timestamping)
> >> + num_channels++;
> >> +
> >> channels = devm_kcalloc(&indio_dev->dev, num_channels,
> >> sizeof(struct iio_chan_spec), GFP_KERNEL);
> >> if (!channels)
> >> @@ -1816,6 +1819,19 @@ static int stm32_adc_chan_of_init(struct iio_dev *indio_dev)
> >> stm32_adc_smpr_init(adc, channels[i].channel, smp);
> >> }
> >>
> >> + if (timestamping) {
> >> + struct iio_chan_spec *timestamp = &channels[scan_index];
> >> +
> >> + timestamp->type = IIO_TIMESTAMP;
> >> + timestamp->channel = -1;
> >> + timestamp->scan_index = scan_index;
> >> + timestamp->scan_type.sign = 's';
> >> + timestamp->scan_type.realbits = 64;
> >> + timestamp->scan_type.storagebits = 64;
> >> +
> >> + scan_index++;
> >> + }
> >> +
> >> indio_dev->num_channels = scan_index;
> >> indio_dev->channels = channels;
> >>
> >> @@ -1875,6 +1891,7 @@ static int stm32_adc_probe(struct platform_device *pdev)
> >> struct device *dev = &pdev->dev;
> >> irqreturn_t (*handler)(int irq, void *p) = NULL;
> >> struct stm32_adc *adc;
> >> + bool timestamping = false;
> >> int ret;
> >>
> >> if (!pdev->dev.of_node)
> >> @@ -1931,16 +1948,22 @@ static int stm32_adc_probe(struct platform_device *pdev)
> >> if (ret < 0)
> >> return ret;
> >>
> >> - ret = stm32_adc_chan_of_init(indio_dev);
> >> - if (ret < 0)
> >> - return ret;
> >> -
> >> ret = stm32_adc_dma_request(dev, indio_dev);
> >> if (ret < 0)
> >> return ret;
> >>
> >> - if (!adc->dma_chan)
> >> + if (!adc->dma_chan) {
> >> + /* For PIO mode only, iio_pollfunc_store_time stores a timestamp
> >> + * in the primary trigger IRQ handler and stm32_adc_trigger_handler
> >> + * runs in the IRQ thread to push out buffer along with timestamp.
> >> + */
> >> handler = &stm32_adc_trigger_handler;
> >> + timestamping = true;
> >> + }
> >> +
> >> + ret = stm32_adc_chan_of_init(indio_dev, timestamping);
> >> + if (ret < 0)
> >> + goto err_dma_disable;
> >>
> >> ret = iio_triggered_buffer_setup(indio_dev,
> >> &iio_pollfunc_store_time, handler,
> >
> >
>
