Re: PATCH? rcu: eliminate rcu_ctrlblk.lock

[Manfred Spraul <manfred@colorfullife.com>]

Oleg Nesterov wrote:

>Hello.
>
>I am trying to understand the rcu implementetion.
>
>I can't understand why rcu_ctrlblk.seqcount is needed.
>It seems to me it can be replaced by a couple of barriers.
>
>  
>

Your patch would add one new corner case:

start: next_pending==1. rcp->cur == 11.
cpu 1: rcu_start_back sets next_pending to 0.
cpu 2: rdp->batch = rcp->cur + 1 [i.e. wait for end of period 12]
cpu 2: notices next_pending == 0, tries to acquire the spinlock [blocks]
cpu 1: rcp->cur++ [i.e. start period 12]
cpu 1: releases the spinlock
cpu 2: gets the spinlock, sets next_pending to 1 and exits.

Now next_pending is 1 [i.e. at the end of grace period 12 grace period 
13 is automatically started], although noone has callbacks waiting for 
period 13.

This is not fatal: the combination is rare, so perhaps we could tolerate 
the race. But on the other hand the sequence locks are outside of the 
hot paths and not much slower than a smp_rmb().

Dipankar - what do you think?

--
    Manfred