demux: make hysteresis-secs respect cache-secs

[N-R-K]

currently hyst_active is only reactivated when the memory limit fills up. but we might stop prefetching due to hitting cache-secs as well, which might occur well before hitting the memory limit. and in those cases, hyst_active never gets reactivated and thus is broken.

this patch activates hysteresis whenever there's nothing more to read or prefetch.

[github-actions]

Download the artifacts for this pull request:

Windows

• mpv-i686-w64-mingw32
• mpv-x86_64-w64-mingw32

[Jules-A]

Need a hysteresis percentage that will cache to a percentage of demuxer-max-bytes or caches-sec, whichever is greater. OR at least a demuxer-hysteresis-bytes as it's hard to get it useful for both high bitrate and low bitrate in it's current state.

[N-R-K]

currently hyst_active is only reactivated when the memory limit fills up

Reading the docs, seems like this was intended. But this isn't intuitive/doesn't make sense since if the goal is to save power by prefetching in bursts then it make sense to do it when cache-secs fills up before max-bytes as well.

[N-R-K]

Reading the docs, seems like this was intended.

But then the docs mention cache-secs and demuxer-readahead-secs as well. That's why I was expecting it to work in the first place, and was surprised to see it still buffer non-stop.

@kerneltoast Since you were the one who added this flag, do you have any comment on all of this?

[N-R-K]

OR at least a demuxer-hysteresis-bytes

This was lightly discussed on the original PR.
Initially I also thought specifying bytes might be better but after trying it out I think this option makes more sense in time than bytes.

[Jules-A]

OR at least a demuxer-hysteresis-bytes

This was lightly discussed on the original PR. Initially I also thought specifying bytes might be better but after trying it out I think this option makes more sense in time than bytes.

I wasn't talking out 1 or the other but the option to specify both (if you ignore the % idea).

[kasper93]

if the goal is to save power by prefetching in bursts then it make sense to do it when cache-secs fills up before max-bytes as well.

Exactly my thoughts, this bothered me too for some time.

Change LGTM and works as advertised. 👍

[Dudemanguy]

Needs a rebase but seems fine.