06.04.2020 20:02, Sowjanya Komatineni пишет: > > On 4/6/20 9:37 AM, Sowjanya Komatineni wrote: >> >> On 4/6/20 9:29 AM, Dmitry Osipenko wrote: >>> External email: Use caution opening links or attachments >>> >>> >>> 06.04.2020 19:12, Sowjanya Komatineni пишет: >>>> On 4/6/20 9:05 AM, Dmitry Osipenko wrote: >>>>> External email: Use caution opening links or attachments >>>>> >>>>> >>>>> 06.04.2020 18:35, Sowjanya Komatineni пишет: >>>>> ... >>>>>>>> + /* wait for syncpt counter to reach frame start event >>>>>>>> threshold */ >>>>>>>> + err = host1x_syncpt_wait(chan->frame_start_sp, thresh, >>>>>>>> + TEGRA_VI_SYNCPT_WAIT_TIMEOUT, &value); >>>>>>>> + if (err) { >>>>>>>> + dev_err(&chan->video.dev, >>>>>>>> + "frame start syncpt timeout: %d\n", err); >>>>>>>> + /* increment syncpoint counter for timedout events */ >>>>>>>> + host1x_syncpt_incr(chan->frame_start_sp); >>>>>>> Why incrementing is done while hardware is still active? >>>>>>> >>>>>>> The sync point's state needs to be completely reset after resetting >>>>>>> hardware. But I don't think that the current upstream host1x driver >>>>>>> supports doing that, it's one of the known-long-standing problems of >>>>>>> the >>>>>>> host1x driver. >>>>>>> >>>>>>> At least the sp->max_val incrementing should be done based on the >>>>>>> actual >>>>>>> syncpoint value and this should be done after resetting hardware. >>>>>> upstream host1x driver don't have API to reset or to equalize max >>>>>> value >>>>>> with min/load value. >>>>>> >>>>>> So to synchronize missed event, incrementing HW syncpt counter. >>>>>> >>>>>> This should not impact as we increment this in case of missed events >>>>>> only. >>>>> It's wrong to touch sync point while hardware is active and it's >>>>> active >>>>> until being reset. >>>>> >>>>> You should re-check the timeout after hw resetting and manually put >>>>> the >>>>> syncpoint counter back into sync only if needed. >>>> There is possibility of timeout to happen any time even during the >>>> capture also and is not related to hw reset. >>>> >>>> Manual synchronization is needed when timeout of any frame events >>>> happen >>>> otherwise all subsequence frames will timeout due to mismatch in event >>>> counters. >>> My point is that hardware is stopped only after being reset, until then >>> you should assume that sync point could be incremented by HW at any >>> time. >>> >>> And if this happens that HW increments sync point after the timeout, >>> then the sync point counter should become out-of-sync in yours case, >>> IIUC. Because host1x_syncpt_incr() doesn't update the cached counter. >> >> We wait for enough time based on frame rate for syncpt increment to >> happen and if it doesn't happen by then definitely its missed event >> and we increment HW syncpoint for this timed event. >> >> cached value gets updated during syncpt wait for subsequent event. >> >> syncpt increment happens for all subsequent frame events during video >> capture. >> > Just to be clear, syncpt max value increment happens first and syncpt > trigger condition is programmed. hw syncpt increment happens based on HW > events. > > Wait time for HW syncpt to reach threshold is tuned to work for all > frame rates. So if increment doesn't happen by then, its definitely > missed event. This is questionable. Technically, speculating about whether the tuned value is good for all possible cases is incorrect thing to do. Although, I guess in practice it should be good enough for the starter and could be improved later on, once the host1x driver will be improved. > In case of missed HW event corresponding to syncpt condition, hw syncpt > increment does not happen and driver increments it on timeout. > > As there is not API to equialize max with min incase of timeout/reset, > incrementing HW syncpt for timed out event. > > syncpt cached value gets updated during syncpt wait when it loads from > HW syncpt. > > As syncpt condition is already triggered, without compensating timeout > events or leaving syncpt max and hw syncpt in non synchronized state for > missed events, subsequent streamings will all timeout even on real events. >