On Fri, 17 Feb 2023 15:43:26 +0100 Mehdi Djait <mehdi.djait.k@xxxxxxxxx> wrote: > On Fri, Feb 17, 2023 at 02:28:28PM +0000, Jonathan Cameron wrote: > > On Fri, 17 Feb 2023 13:59:16 +0200 > > Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote: > > > > > On 2/17/23 13:43, Jonathan Cameron wrote: > > > > On Fri, 17 Feb 2023 07:56:22 +0200 > > > > Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote: > > > > > > > >> Hi Mehdi, > > > >> > > > >> On 2/16/23 22:22, Mehdi Djait wrote: >>> And here are the relevant steps after an IRQ occurs : > > > >>> 1. IRQ context --> kx022a_irq_handler() --> gets the current timestamp > > > >>> with "data->timestamp = iio_get_time_ns(idev);" and returns > > > >>> IRQ_WAKE_THREAD > > > >>> > > > >>> 2. kx022a_irq_thread_handler() -> checks that the trigger is enabled > > > >>> --> iio_trigger_poll_chained() --> handle_nested_irq(): which will only > > > >>> call the bottom half of the pollfuncs > > > >> > > > >> I don't get the kx022a at my hands until next week to test this, but it > > > >> seems to me your reasoning is right. iio_pollfunc_store_time() is > > > >> probably not called. I just wonder why I didn't see zero timestamps when > > > >> testing this. (OTOH, I had somewhat peculiar IRQ handling at first - > > > >> maybe I broke this along the way). > > > > > > > > This is a common problem. So far we've always solved it in the driver > > > > by using the pf->timestamp only if it's been set - otherwise fallback > > > > to grabbing a new one to pass into iio_push_to_buffer_with_timestamp() > > > > in the threaded handler. > > > > > > > > It might be possible to solve in a generic fashion but it's a bit > > > > fiddly so I don't think anyone has ever looked at it. > > > > > > I agree it's "fiddly" :) I played with a though of conditionally adding > > > the timestamp in the iio_trigger_poll_chained() if the timestamp is zero > > > there. This, however, would require clearing the timestamp when it is > > > read - which gets "fiddly" soon. Hence I just suggested adding a note in > > > kerneldoc. > > > > > > >> > > > >>> Question 2: If the change proposed in question 1 is wrong, would this > > > >>> one be better iio_push_to_buffers_with_timestamp(idev, data->buffer, > > > >>> iio_get_time_ns(idev)). There is some delay between the IRQ occuring > > > >>> and trigger_handler being called but that is better than getting all 0 > > > >>> timestamps like suggested in [2] > > > >> > > > >> Please, use the data->timestamp as you suggested. > > > > > > > > I'd suggest a bit of both. If you have a timestamp from the irq handler > > > > use it. If it's not available then grab one locally in the threaded handler. > > > > > > Hm. I don't think we will end up in the kx022a threaded handler so that > > > the data->timestamp is not populated in the IRQ handler. I am _far_ from > > > an IIO expert - but I guess the only way would be that some other > > > trigger invoked the threaded handler(?) Shouldn't the > > > kx022a_validate_trigger() prevent this? > > Ah. I'd missed this one restricted what triggers could be used. > > We'll have to pay attention to this if that particular condition is ever > > relaxed. > > > > > > Please, follow Jonathan's guidance if he does not tell othervice. You > > > clearly should not trust a random guy who obviously does not know how to > > > write these drivers in the first place XD > > > > You were right here :) > > So should I send a patch with data->timestamp as I suggested ? I think so. > > And should I write some documentaion to highlight the > difference between iio_trigger_poll and iio_trigger_poll_chained, i.e., > where the functions expect to be called ? > Something similar to the /kernel/irq/irqdesc.c Sure. The chained naming is inconsistent, but I never got around to fixing it. Probably should be something like iio_trigger_poll_nested() Comment on expectations is good. > > > > > > > > > >> > > > >>> I hope that I'm understating this correctly or at least not totally > > > >>> off :) If yes, I will send a patch. > > > >> > > > >> Thanks Mehdi! I think this was a great catch! Maybe - while at it - you > > > >> could also send a patch adding a small kerneldoc to the > > > >> iio_trigger_poll_chained() mentioning this particular issue. Yes, I > > > >> guess it should be obvious just by reading the function name *_chained() > > > >> - but I did fall on this trap (and according to your reference [2] so > > > >> has someone else). > > > >> > > > >>> [1] https://lore.kernel.org/linux-iio/4FDB33CD.2090805@xxxxxxxxxx/ > > > >>> [2] https://lore.kernel.org/linux-iio/20201205182659.7cd23d5b@archlinux/ > > > >>> [3] https://lore.kernel.org/linux-iio/20220126191606.00003f37@xxxxxxxxxx/ > > > >> > > > >> Yours, > > > >> -- Matti > > > >> > > > > > > > > >
