Re: [thiscpuops upgrade 10/10] Lockless (and preemptless) fastpaths for slub

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* Christoph Lameter (cl@linux.com) wrote:

[...]

> @@ -1737,23 +1770,53 @@ static __always_inline void *slab_alloc(
>  {
>  	void **object;
>  	struct kmem_cache_cpu *c;
> -	unsigned long flags;
> +	unsigned long tid;
>  
>  	if (slab_pre_alloc_hook(s, gfpflags))
>  		return NULL;
>  
> -	local_irq_save(flags);
> +redo:
> +	/*
> +	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
> +	 * enabled. We may switch back and forth between cpus while
> +	 * reading from one cpu area. That does not matter as long
> +	 * as we end up on the original cpu again when doing the cmpxchg.
> +	 */
>  	c = __this_cpu_ptr(s->cpu_slab);
> +
> +	/*
> +	 * The transaction ids are globally unique per cpu and per operation on
> +	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
> +	 * occurs on the right processor and that there was no operation on the
> +	 * linked list in between.
> +	 */

There seems to be some voodoo magic I don't understand here. I'm curious to see
what happens if we have:

CPU A                                                  CPU B
slab_alloc()
  c = __this_cpu_ptr(s->cpu_slab);
  tid = c->tid
  thread migrated to CPU B

slab_alloc()
  c = __this_cpu_ptr(s->cpu_slab);
  tid = c->tid
  ...                                                  ...
  irqsafe_cmpxchg_double
    - expect tid, on CPU A, success
                                                       migrate back to CPU A
  irqsafe_cmpxchg_double
    - expect (same) tid, on CPU A, success

So either there is a crucially important point I am missing, or the transaction
ID does not seem to be truly unique due to migration.

Thanks,

Mathieu


> +	tid = c->tid;
> +	barrier();
> +
>  	object = c->freelist;
> -	if (unlikely(!object || !node_match(c, node)))
> +	if (unlikely(!object || !node_match(c, c->node)))
>  
> -		object = __slab_alloc(s, gfpflags, node, addr, c);
> +		object = __slab_alloc(s, gfpflags, c->node, addr);
>  
>  	else {
> -		c->freelist = get_freepointer(s, object);
> +		/*
> +		 * The cmpxchg will only match if there was not additonal
> +		 * operation and if we are on the right processor.
> +		 */
> +		if (unlikely(!irqsafe_cmpxchg_double(&s->cpu_slab->freelist, object, tid,
> +				get_freepointer(s, object), next_tid(tid)))) {


-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com