Re: [PATCH 1/2] net: reimplement completion_queue as a FIFO queue

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 26 avril 2010 à 13:44 +0800, Changli Gao a écrit :
> On Mon, Apr 26, 2010 at 1:14 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Le lundi 26 avril 2010 à 11:20 +0800, Changli Gao a écrit :
> >> reimplement completion_queue as a FIFO queue.
> >>
> >> As slab allocator always does its best to return the latest unused objects, we'd
> >> better release skb in order, this patch reimplement completion_queue as a FIFO
> >> queue instead of the old LIFO queue.
> >
> >
> > 1) New devices dont use completion queue.
> 
> It is good enough reason for rejection.
> 

I said, this part of the kernel is not used today.

> >
> > 2) Hot objects are the last enqueued.
> >   If many objects were queued, the old one are not hot anymore.
> >
> >   Using FIFO will give more cache misses, when walking the chain
> > (skb->next)
> >
> 
> I meaned that slab allocator maintains objects in the LIFO manner, and
> if we call kmem_cache_alloc(cache) after a kmem_cache_free(cache)
> call, the last object freed by kmem_cache_free(cache) will be
> returned, as this object are more likely in cache and hot. If we don't
> realse skbs in FIFO manner, the last and hot one will not been
> returned at first.
> 

But your patch doesnt change this behavior. This is irrelevant.

LIFO is the slub behavior, for the exact reason its better for cache
reuse. Same for completion queue.

I repeat : 
- slub dont touch objects in normal situations.
- LIFO is better for caches.
- LIFO is faster (one pointer for the queue head)