Re: [PATCH v2 1/3] lib/list_batch: A simple list insertion/deletion batching facility

[Waiman Long]

On 01/31/2016 07:47 PM, Dave Chinner wrote:
> On Fri, Jan 29, 2016 at 02:30:44PM -0500, Waiman Long wrote:
> > Linked list insertion or deletion under lock is a very common activity
> > in the Linux kernel. If this is the only activity under lock, the
> > locking overhead can be pretty large compared with the actual time
> > spent on the insertion or deletion operation itself especially on a
> > large system with many CPUs.
> >
> > This patch introduces a simple list insertion/deletion batching
> > facility where a group of list insertion and deletion operations are
> > grouped together in a single batch under lock. This can reduce the
> > locking overhead and improve overall system performance.
> >
> > The fast path of this batching facility will be similar in performance
> > to the "lock; listop; unlock;" sequence of the existing code. If
> > the lock is not available, it will enter slowpath where the batching
> > happens.
> >
> > A new config option LIST_BATCHING is added so that we can control on
> > which architecture do we want to have this facility enabled.
> >
> > Signed-off-by: Waiman Long<Waiman.Long@xxxxxxx>
> ....
> > +#ifdef CONFIG_LIST_BATCHING
> > +
> > +extern void do_list_batch_slowpath(spinlock_t *lock, enum list_batch_cmd cmd,
> > +				   struct list_batch *batch,
> > +				   struct list_head *entry);
> > +
> > +/*
> > + * The caller is expected to pass in a constant cmd parameter. As a
> > + * result, most of unneeded code in the switch statement of _list_batch_cmd()
> > + * will be optimized away. This should make the fast path almost as fast
> > + * as the "lock; listop; unlock;" sequence it replaces.
> > + */
> This strikes me as needlessly complex. Simple inline functions are
> much easier to read and verify correct, and we don't have to rely on
> the compiler to optimise out dead code:
>
> static inline void list_batch_add(struct list_head *entry,
> 				  struct list_batch *batch)
> {
> 	if (!spin_trylock(&batch->lock))
> 		return do_list_batch_slowpath(entry, batch, lb_cmd_add);
>
> 	list_add(entry,&batch->list)
> 	spin_unlock(&batch->lock);
> }

Will do so.

> > +#include<linux/list_batch.h>
> > +
> > +/*
> > + * List processing batch size = 128
> > + *
> > + * The batch size shouldn't be too large. Otherwise, it will be too unfair
> > + * to the task doing the batch processing. It shouldn't be too small neither
> > + * as the performance benefit will be reduced.
> > + */
> > +#define LB_BATCH_SIZE	(1<<  7)
> Ok, so arbitrary operations are going to see longer delays when they
> are selected as the batch processor. I'm not sure I really like this
> idea, as it will be the first in the queue that sees contention
> that takes the delay which reduces the fairness of the operations.
> i.e. the spinlock uses fair queuing, but now we can be grossly unfair
> the to the first spinner...

That is certainly true. It is well known that a bit of unfairness can 
often improve overall system performance. Interrupt handling, for 
example, is also unfair to the process currently running on the CPU. The 
amount of unfairness is controlled by the batch size parameter. Maybe we 
can make this parameter a read-mostly constant set up at boot time which 
has a value depending on the # of CPUs in a system so that smaller 
system can have a smaller batch size. That will reduce the unfairness, 
at least in smaller systems.

> > +	/*
> > +	 * We rely on the implictit memory barrier of xchg() to make sure
> > +	 * that node initialization will be done before its content is being
> > +	 * accessed by other CPUs.
> > +	 */
> > +	prev = xchg(&batch->tail,&node);
> > +
> > +	if (prev) {
> > +		WRITE_ONCE(prev->next,&node);
> > +		while (READ_ONCE(node.state) == lb_state_waiting)
> > +			cpu_relax();
> > +		if (node.state == lb_state_done)
> > +			return;
> So we spin waiting for the batch processor to process the
> list, or
>
> > +		WARN_ON(node.state != lb_state_batch);
> tell us we are not the batch processor.
>
> So, effectively, the reduction in runtime is due to the fact the
> list operations spin on their own cache line rather than the spin
> lock cacheline until they have been processed and/or made the batch
> processor?

Yes, that can a major part of it.

> > +	}
> > +
> > +	/*
> > +	 * We are now the queue head, we should acquire the lock and
> > +	 * process a batch of qnodes.
> > +	 */
> > +	loop = LB_BATCH_SIZE;
> > +	next =&node;
> > +	spin_lock(lock);
> > +
> > +do_list_again:
> > +	do {
> While we are batch processing, all operations will fail the
> trylock and add themselves to the tail of the queue, and spin on
> their own cacheline at that point. So it doesn't reduce the amount
> of spinning, just removes the cacheline contention that slows the
> spinning.
>
> Hmmm - there's another point of unfairness - when switching batch
> processors, other add/delete operations can get the list lock and
> perform their operations directly, thereby jumping the batch
> queue....

That is true too.

> So at what point does simply replacing the list_head with a list_lru
> become more efficient than this batch processing (i.e.
> https://lkml.org/lkml/2015/3/10/660)?  The list_lru isn't a great
> fit for the inode list (doesn't need any of the special LRU/memcg
> stuff https://lkml.org/lkml/2015/3/16/261) but it will tell us if,
> like Ingo suggested, moving more towards a generic per-cpu list
> would provide better overall performance...

I will take a look at the list_lru patch to see if that help. As for the 
per-cpu list, I tried that and it didn't quite work out.

Thanks,
Longman