Messages in this thread | | | Date | Wed, 6 Aug 2008 20:18:06 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH tip/core/rcu] classic RCU locking and memory-barrier cleanups |
| |
On Wed, Aug 06, 2008 at 07:30:13AM +0200, Manfred Spraul wrote: > Hi Paul, > > Paul E. McKenney wrote: >> This patch is in preparation for moving to a hierarchical >> algorithm to allow the very large SMP machines -- requested by some >> people at OLS, and there seem to have been a few recent patches in the >> 4096-CPU direction as well. > > I thought about hierarchical RCU, but I never found the time to implement > it. > Do you have a concept in mind?
Actually, you did submit a patch for a two-level hierarchy some years back:
http://marc.theaimsgroup.com/?l=linux-kernel&m=108546384711797&w=2
I am looking to allow multiple levels to accommodate 4096 CPUs, which pushes me towards locking on the nodes in the hierarchy. I have a roughed-out design that (I hope!) avoids deadlock and that allows adapting to machine topology. I am also trying to minimize the amount of arch-specific code needed to construct the hierarchy -- hopefully just a pair of config parameters.
More as it starts working...
> Right now, I try to understand the current code first - and some of it > doesn't make much sense. > > There are three per-cpu lists: > ->nxt > ->cur > ->done. > > Obviously, there must be a quiescent state between cur and done. > But why does the code require a quiescent state between nxt and cur? > I think that's superflous. The only thing that is required is that all cpus > have moved their callbacks from nxt to cur. That doesn't need a quiescent > state, this operation could be done in hard interrupt as well.
The deal is that we have to put incoming callbacks somewhere while the batch in ->cur waits for an RCU grace period. That somewhere is ->nxt. So to be painfully pedantic, the callbacks in ->nxt are not waiting for an RCU grace period. Instead, they are waiting for the callbacks in ->cur to get out of the way.
> Thus I think this should work: > > 1) A callback is inserted into ->nxt.
Yep.
> 2) As soon as too many objects are sitting in the ->nxt lists, a new rcu > cycle is started.
Yep, call_rcu() and friends now do this. (In response to denial of services attacks some years back.)
> 3) As soon as a cpu sees that a new rcu cycle is started, it moves it's > callbacks from ->nxt to ->cur. No checks for hard_irq_count & friends > necessary. Especially: same rule for _bh and normal.
Yep. The checks for hard_irq_count are instead intended to determine if this CPU is already in a quiescent state for the newly started RCU grace period. As long as we took the scheduling clock interrupt, we might as well get our money's worth, right?
> 4) As soon as all cpus have moved their lists from ->nxt to ->cur, the real > grace period is started.
Jiangshan took a slightly different approach to handling this situation, but yes, more or less. The trick is that the processing in (4) for ->nxt is overlapped with the processing in (5) for ->cur.
> 5) As soon as all cpus passed a quiescent state (i.e.: now with tests for > hard_irq_count, different rules for _bh and normal), the list is moved from > ->cur to ->completed. Once in completed, they can be destroyed by > performing the callbacks.
To ->done rather than ->completed, but yes.
> What do you think? would that work? It doesn't make much sense that step 3) > tests for a quiescent state.
The trick is that the work for grace period n and grace period n+1 are overlapped.
> Step 2) could depend memory pressure.
Yep.
> Step 3) and 4) could be accelerated by force_quiescent_state(), if the > memory pressure is too high.
Yep -- though we haven't done this except on paper.
Thanx, Paul
> -- > Manfred > -> nxt >
| |