Re: [PATCH 4/5] rcu: join rcu_ctrlblk and rcu_state

[Manfred Spraul <manfred@colorfullife.com>]

[I haven't read the diff, just a short comment]

Dipankar Sarma wrote:

>rcu_state came over from Manfred's RCU_HUGE patch IIRC. I don't
>think it is necessary to allocate rcu_state separately in the
>current mainline RCU code. So, the patch looks OK to me, but
>Manfred might see something that I am not seeing.
>
>  
>
The two-level rcu code was never merged, I still plan to clean it up.

But the idea of splitting the control block and the state is used in the 
current code:
- __rcu_pending() is the hot path, it only performs a read access to 
rcu_ctrlblk.
- write accesses to the rcu_ctrlblk are really rare, they only happen 
when a new batch is started. Especially: independant from the number of 
cpus.

Write access to the rcu_state are common:
- each cpu must write once in each cycle to update it's cpu mask.
- The last cpu then completes the quiescent cycle.

The idea is that rcu_state is more or less write-only and rcu_state is 
read-only. Theoretically, rcu_state could be shared in all cpus caches, 
and there will be only one invalidate when a new batch is started. Thus 
no cacheline trashing due to rcu_pending calls.
I think it would be safer to keep the two state counters in a separate 
cacheline from the spinlock and the cpu mask, but I don't have any hard 
numbers. IIRC the problems with the large SGI systems disappered, and 
everyone was happy. No real benchmark comparisons were made.

--
    Manfred