Received-SPF: pass (google.com: domain of linux-kernel+bounces-81162-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <20240220021804.9541-1-shijie@os.amperecomputing.com>
 <CANn89iJoHDzfYfhcwVvR4m7DiVG-UfFNqm+D1WD-2wjOttk6ew@mail.gmail.com>
 <bea860f8-a196-4dff-a655-4da920e2ebfa@amperemail.onmicrosoft.com>
 <CANn89i+1uMAL_025rNc3C1Ut-E5S8Nat6KhKEzcFeC1xxcFWaA@mail.gmail.com>
 <c2bd73b6-b21f-4ad8-a176-eec677bc6cf3@amperemail.onmicrosoft.com>
 <CANn89i+Cr1Tbdxqy6fB-sOLca+AHFc-3-0xGktVUsQFFMVsC0A@mail.gmail.com> <bd0a9d66-783d-4936-a5b0-cd4082704137@kernel.org>
In-Reply-To: <bd0a9d66-783d-4936-a5b0-cd4082704137@kernel.org>
From: Eric Dumazet <edumazet@google.com>
Date: Mon, 26 Feb 2024 11:29:07 +0100
Message-ID: <CANn89iKJtLZFjY_kYhV5NcgKiAkhi_stftvai1dCwQLMOLea6g@mail.gmail.com>
Subject: Re: [PATCH] net: skbuff: allocate the fclone in the current NUMA node
To: Jesper Dangaard Brouer <hawk@kernel.org>
Cc: aleksander.lobakin@intel.com, 
	Shijie Huang <shijie@amperemail.onmicrosoft.com>, 
	Huang Shijie <shijie@os.amperecomputing.com>, kuba@kernel.org, 
	patches@amperecomputing.com, davem@davemloft.net, horms@kernel.org, 
	ast@kernel.org, dhowells@redhat.com, linyunsheng@huawei.com, 
	linux-kernel@vger.kernel.org, netdev@vger.kernel.org, 
	cl@os.amperecomputing.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Mon, Feb 26, 2024 at 11:18=E2=80=AFAM Jesper Dangaard Brouer <hawk@kerne=
l.org> wrote:
>
>
>
> On 24/02/2024 20.07, Eric Dumazet wrote:
> > On Tue, Feb 20, 2024 at 9:37=E2=80=AFAM Shijie Huang
> > <shijie@amperemail.onmicrosoft.com> wrote:
> >>
> >>
> >> =E5=9C=A8 2024/2/20 16:17, Eric Dumazet =E5=86=99=E9=81=93:
> >>> On Tue, Feb 20, 2024 at 7:26=E2=80=AFAM Shijie Huang
> >>> <shijie@amperemail.onmicrosoft.com> wrote:
> >>>>
> >>>> =E5=9C=A8 2024/2/20 13:32, Eric Dumazet =E5=86=99=E9=81=93:
> >>>>> On Tue, Feb 20, 2024 at 3:18=E2=80=AFAM Huang Shijie
> >>>>> <shijie@os.amperecomputing.com> wrote:
> >>>>>> The current code passes NUMA_NO_NODE to __alloc_skb(), we found
> >>>>>> it may creates fclone SKB in remote NUMA node.
> >>>>> This is intended (WAI)
> >>>> Okay. thanks a lot.
> >>>>
> >>>> It seems I should fix the issue in other code, not the networking.
> >>>>
> >>>>> What about the NUMA policies of the current thread ?
> >>>> We use "numactl -m 0" for memcached, the NUMA policy should allocate
> >>>> fclone in
> >>>>
> >>>> node 0, but we can see many fclones were allocated in node 1.
> >>>>
> >>>> We have enough memory to allocate these fclones in node 0.
> >>>>
> >>>>> Has NUMA_NO_NODE behavior changed recently?
> >>>> I guess not.
> >>>>> What means : "it may creates" ? Please be more specific.
> >>>> When we use the memcached for testing in NUMA, there are maybe 20% ~=
 30%
> >>>> fclones were allocated in
> >>>>
> >>>> remote NUMA node.
> >>> Interesting, how was it measured exactly ?
> >>
> >> I created a private patch to record the status for each fclone allocat=
ion.
> >>
> >>
> >>> Are you using SLUB or SLAB ?
> >>
> >> I think I use SLUB. (CONFIG_SLUB=3Dy,
> >> CONFIG_SLAB_MERGE_DEFAULT=3Dy,CONFIG_SLUB_CPU_PARTIAL=3Dy)
> >>
> >
> > A similar issue comes from tx_action() calling __napi_kfree_skb() on
> > arbitrary skbs
> > including ones that were allocated on a different NUMA node.
> >
> > This pollutes per-cpu caches with not optimally placed sk_buff :/
> >
> > Although this should not impact fclones, __napi_kfree_skb() only ?
> >
> > commit 15fad714be86eab13e7568fecaf475b2a9730d3e
> > Author: Jesper Dangaard Brouer <brouer@redhat.com>
> > Date:   Mon Feb 8 13:15:04 2016 +0100
> >
> >      net: bulk free SKBs that were delay free'ed due to IRQ context
> >
> > What about :
> >
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index c588808be77f563c429eb4a2eaee5c8062d99582..63165138c6f690e14520f11=
e32dc16f2845abad4
> > 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -5162,11 +5162,7 @@ static __latent_entropy void
> > net_tx_action(struct softirq_action *h)
> >                                  trace_kfree_skb(skb, net_tx_action,
> >                                                  get_kfree_skb_cb(skb)-=
>reason);
> >
> > -                       if (skb->fclone !=3D SKB_FCLONE_UNAVAILABLE)
> > -                               __kfree_skb(skb);
> > -                       else
> > -                               __napi_kfree_skb(skb,
> > -                                                get_kfree_skb_cb(skb)-=
>reason);
>
> Yes, I think it makes sense to avoid calling __napi_kfree_skb here.
> The __napi_kfree_skb call will cache SKB slub-allocation (but "release"
> data) on a per CPU napi_alloc_cache (see code napi_skb_cache_put()).
> In net_tx_action() there is a chance this could originate from another
> CPU or even NUMA node.  I notice this is only for SKBs on the
> softnet_data->completion_queue, which have a high chance of being cache
> cold.  My patch 15fad714be86e only made sense when we bulk freed these
> SKBs, but after Olek's changes to cache freed SKBs, then this shouldn't
> be calling __napi_kfree_skb() (previously named __kfree_skb_defer).
>
> I support this RFC patch from Eric.
>
> Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>

Note that this should not matter for most NIC, because their drivers
perform TX completion from NAPI context, we do not hit this path.

It seems that switching to SLUB instead of SLAB has increased the chances
of getting memory from another node.

We probably need to investigate.