Re: [RFC][PATCH v2 08/11] context_tracking,rcu: Replace RCU dynticks counter with context_tracking

[Peter Zijlstra]

On Wed, Sep 29, 2021 at 11:37:01AM -0700, Paul E. McKenney wrote:
> On Wed, Sep 29, 2021 at 05:17:31PM +0200, Peter Zijlstra wrote:
> The CT_SEQ_WORK bit means neither idle nor nohz_full user, correct?

That bit is indeed independent, it can get set remotely by cmpxchg when
in user/eqs state and will be tested and (eventually) cleared when
leaving user/eqs state.

> So let's see if I intuited the decoder ring, where "kernel" means that
> portion of the kernel that is non-noinstr...
> 
> CT_SEQ_WORK	CT_SEQ_NMI	CT_SEQ_USER	Description?
> 0		0		0		Idle or non-nohz_full user
> 0		0		1		nohz_full user
> 0		1		0		NMI from 0,0,0
> 0		1		1		NMI from 0,0,1
> 1		0		0		Non-idle kernel
> 1		0		1		Cannot happen?
> 1		1		0		NMI from 1,0,1
> 1		1		1		NMI from cannot happen
> 
> And of course if a state cannot happen, Murphy says that you will take
> an NMI in that state.

Urgh, you have the bits the 'wrong' way around :-)

MSB     3210
|-------||||

Where [MSB-3] is the actual sequence number, [21] is the state and [0]
is the work-pending bit.

The table for [21] is like:

USER NMI
 0    0		kernel
 0    1		kernel took nmi
 1    0		user
 1    1		user took nmi

So effectively EQS is 10 only, the other 3 states are kernel.