Re: [PATCH] ipc/msg: Implement lockless pipelined wakeups

From: Davidlohr Bueso
Date: Sat Oct 31 2015 - 14:55:17 EST


On Sat, 31 Oct 2015, George Spelvin wrote:

Don't we need to keep that NULL init? I might be missing something.

I wondered the same thing, but on reading it, the cleanup is that he's
gotten rid of the need for the entire thing. Previously, there was a
mechanism for detecting "wakeup not quite finished" that used a NULL
value, but it's no longer needed.

The resultant busy-waiting on the part of the woken-up task was the
entire problem this patch aims to fix. So it gets rid of a whole lot
of code and barriers. And, as you noticed, the comments explaining them.

As the old code explained, the issue is that a task may exit as
soon as r_msg is set, so the wakeup procedure has to be:
- Ensure r_msg is set to NULL (special-case flag)
- Do the wake up
- Set r_msg to the final value

The woken-up task has to spin as long as r_msg is NULL. Ick.

I agree this is a nice cleanup. It is similar to what we did
with posix mqueues, but this one gets rid of even more code afaict.

However, a wake_q keeps a reference to a task, so exiting is
not a danger. As long as wake_q_add precedes setting r_msg,
all is well.

Yes, and this confirms that we still rely on the implicit barrier
from the cmpxchg as tglx mentioned. As such, we also need to keep
the pairing when reading 'r_msg' in do_msgrcv(), instead of dropping
the comments.

Thanks,
Davidlohr
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/