Deadlock due to too many fake leave timeouts

[s5z]

Deadlock due to too many fake leave timeouts

[s5z]

That's a good point. You can just let all the SYS_futex calls skip the fake leave mechanism, and the futex wait/wake matching code will ensure they stay synchronous. The reason that's not done right now is that both mechanisms were developed independently.

[gaomy3832]

I would like to compose a pull request for this issue. My current solution is when timeout is detected and we decide to blacklist a fake leave, instead of only do one leave on the current one, do real leave on all fake-leave threads that wait on the same PC of the same process. This could significantly accelerate the progress when many threads are used, as they will likely wait on the same sync primitive. Now we don't need to wait for timeout for each of them.

This modification has minimum change to the current code in schedule.cpp/h, so it will be an easy fix. The most correct thing, however, will be to skip futex altogether.

If you think this is useful, let me know and I will make a pull request.

[Joyguo]

I think it will be very useful as the runtime can be shortened. Could you share your patch? I want to shorten the runtime of a large application. Thank you.

[s5z]

Mingyu, go ahead & submit the pull request