Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210111182655.12159-1-alobakin@pm.me>
In-Reply-To: <20210111182655.12159-1-alobakin@pm.me>
From:   Eric Dumazet <edumazet@google.com>
Date:   Tue, 12 Jan 2021 09:20:39 +0100
Message-ID: <CANn89iKceTG_Mm4RrF+WVg-EEoFBD48gwpWX=GQiNdNnj2R8+A@mail.gmail.com>
Subject: Re: [PATCH net-next 0/5] skbuff: introduce skbuff_heads bulking and reusing
To:     Alexander Lobakin <alobakin@pm.me>
Cc:     "David S. Miller" <davem@davemloft.net>,
        Jakub Kicinski <kuba@kernel.org>,
        Edward Cree <ecree@solarflare.com>,
        Jonathan Lemon <jonathan.lemon@gmail.com>,
        Willem de Bruijn <willemb@google.com>,
        Miaohe Lin <linmiaohe@huawei.com>,
        Steffen Klassert <steffen.klassert@secunet.com>,
        Guillaume Nault <gnault@redhat.com>,
        Yadu Kishore <kyk.segfault@gmail.com>,
        Al Viro <viro@zeniv.linux.org.uk>,
        netdev <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jan 11, 2021 at 7:27 PM Alexander Lobakin <alobakin@pm.me> wrote:
>
> Inspired by cpu_map_kthread_run() and _kfree_skb_defer() logics.
>
> Currently, all sorts of skb allocation always do allocate
> skbuff_heads one by one via kmem_cache_alloc().
> On the other hand, we have percpu napi_alloc_cache to store
> skbuff_heads queued up for freeing and flush them by bulks.
>
> We can use this struct to cache and bulk not only freeing, but also
> allocation of new skbuff_heads, as well as to reuse cached-to-free
> heads instead of allocating the new ones.
> As accessing napi_alloc_cache implies NAPI softirq context, do this
> only for __napi_alloc_skb() and its derivatives (napi_alloc_skb()
> and napi_get_frags()). The rough amount of their call sites are 69,
> which is quite a number.
>
> iperf3 showed a nice bump from 910 to 935 Mbits while performing
> UDP VLAN NAT on 1.2 GHz MIPS board. The boost is likely to be
> way bigger on more powerful hosts and NICs with tens of Mpps.

What is the latency cost of these bulk allocations, and for TCP traffic
on which GRO is the norm ?

Adding caches is increasing cache foot print when the cache is populated.

I wonder if your iperf3 numbers are simply wrong because of lack of
GRO in this UDP VLAN NAT case.

We are adding a log of additional code, thus icache pressure, that
iperf3 tests can not really measure.

Most linus devices simply handle one packet at a time (one packet per interrupt)