Re: [PATCH] rcu: Report a quiescent state when it's exactly in the state

From: Byungchul Park
Date: Sun May 13 2018 - 23:13:24 EST




On 2018-05-13 ìì 2:26, Steven Rostedt wrote:
On Sat, 12 May 2018 07:41:19 -0700
"Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> wrote:

Don't get me wrong, this discussion was quite useful to me. We probably
need to at least change the comments, and perhaps the code as well. But
I agree that we need input from Peter and Steven to make much more forward
progress.

It's the weekend so I skimmed more than read this thread, but I will
just add this.

The table Joel posted is interesting, and perhaps we should keep things
consistent with that. But that said, with respect to task-RCU, as
nothing on a trampoline should ever call cond_resched() (and perhaps I
should add code in lockdep that verifies this), we just want a
quiescent state that tells us that the task has left the trampoline. A
cond_resched() should be one of those points that does.

It really has nothing to do with scheduling or preemption. The issue is
that if a task is on a trampoline and gets preempted, there's no
knowing when it is off that trampoline where we can free it. We need to
have places in the kernel that we know is a quiescent state to move
task-RCU forward. cond_resched() seems to be one of them. schedule
itself can not be, because it can be called from an interrupt preempting
a task while it is on the trampoline.

Exactly. I think Steven explained how we should consider them exactly.

-- Steve
--
Thanks,
Byungchul