lkml.org 
[lkml]   [2021]   [Feb]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [v3 net-next 08/10] skbuff: reuse NAPI skb cache on allocation path (__build_skb())
On Wed, 10 Feb 2021 12:25:04 +0000
Alexander Lobakin <alobakin@pm.me> wrote:

> From: Paolo Abeni <pabeni@redhat.com>
> Date: Wed, 10 Feb 2021 11:21:06 +0100
>
> > I'm sorry for the late feedback, I could not step-in before.
> >
> > Also adding Jesper for awareness, as he introduced the bulk free
> > infrastructure.

Thanks (and Alexander Duyck also did part of the work while at Red Hat).

In my initial versions of my patchsets I actually also had reuse of the
SKBs that were defer freed during NAPI context. But I dropped that
part because it was getting nitpicked and the merge window was getting
close, so I ended up dropping that part.



> > On Tue, 2021-02-09 at 20:48 +0000, Alexander Lobakin wrote:
> > > @@ -231,7 +256,7 @@ struct sk_buff *__build_skb(void *data, unsigned int frag_size)
> > > */
> > > struct sk_buff *build_skb(void *data, unsigned int frag_size)
> > > {
> > > - struct sk_buff *skb = __build_skb(data, frag_size);
> > > + struct sk_buff *skb = __build_skb(data, frag_size, true);
> >
> > I must admit I'm a bit scared of this. There are several high speed
> > device drivers that will move to bulk allocation, and we don't have any
> > performance figure for them.
> >
> > In my experience with (low end) MIPS board, cache misses cost tend to
> > be much less visible there compared to reasonably recent server H/W,
> > because the CPU/memory access time difference is much lower.
> >
> > When moving to higher end H/W the performance gain you measured could
> > be completely countered by less optimal cache usage.
> >
> > I fear also latency spikes - I'm unsure if a 32 skbs allocation vs a
> > single skb would be visible e.g. in a round-robin test. Generally
> > speaking bulk allocating 32 skbs looks a bit too much. IIRC, when
> > Edward added listification to GRO, he did several measures with
> > different list size and found 8 to be the optimal value (for the tested
> > workload). Above such number the list become too big and the pressure
> > on the cache outweighted the bulking benefits.
>
> I can change to logics the way so it would allocate the first 8.
> I think I've already seen this batch value somewhere in XDP code,
> so this might be a balanced one.

(Speaking about SLUB code): Bulk ALLOC side disables interrupts, and
can call slow path (___slab_alloc), which is bad for latency sensitive
workloads. This I don't recommend large bulk ALLOCATIONS.

> Regarding bulk-freeing: can the batch size make sense when freeing
> or it's okay to wipe 32 (currently 64 in baseline) in a row?

(Speaking about SLUB code): You can bulk FREE large amount of object
without hurting latency sensitive workloads, because it doesn't disable
interrupts (I'm quite proud that this was possible).


> > Perhaps giving the device drivers the ability to opt-in on this infra
> > via a new helper - as done back then with napi_consume_skb() - would
> > make this change safer?
>
> That's actually a very nice idea. There's only a little in the code
> to change to introduce an ability to take heads from the cache
> optionally. This way developers could switch to it when needed.

Well, I actually disagree that this should be hidden behind a switch
for drivers to enable, as this will take forever to get proper enabled.


> Thanks for the suggestions! I'll definitely absorb them into the code
> and give it a test.
>
> > > @@ -838,31 +863,31 @@ void __consume_stateless_skb(struct sk_buff *skb)
> > > kfree_skbmem(skb);
> > > }
> > >
> > > -static inline void _kfree_skb_defer(struct sk_buff *skb)
> > > +static void napi_skb_cache_put(struct sk_buff *skb)
> > > {
> > > struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
> > > + u32 i;
> > >
> > > /* drop skb->head and call any destructors for packet */
> > > skb_release_all(skb);
> > >
> > > - /* record skb to CPU local list */
> > > + kasan_poison_object_data(skbuff_head_cache, skb);
> > > nc->skb_cache[nc->skb_count++] = skb;
> > >
> > > -#ifdef CONFIG_SLUB
> > > - /* SLUB writes into objects when freeing */
> > > - prefetchw(skb);
> > > -#endif
> >
> > It looks like this chunk has been lost. Is that intentional?
>
> Yep. This prefetchw() assumed that skbuff_heads will be wiped
> immediately or at the end of network softirq. Reusing this cache
> means that heads can be reused later or may be kept in a cache for
> some time, so prefetching makes no sense anymore.

I agree with this statement, the prefetchw() is no-longer needed.

--
Best regards,
Jesper Dangaard Brouer
MSc.CS, Principal Kernel Engineer at Red Hat
LinkedIn: http://www.linkedin.com/in/brouer

\
 
 \ /
  Last update: 2021-02-10 18:29    [W:0.051 / U:21.672 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site